Research and Quantitative Analysis on Dynamic Risk Assessment of Intelligent Connected Vehicles

Comments 1: The abstract is informative but somewhat dense. If you have the possibility of editing consider separating methodological contributions from performance results for clarity.

Lines 23-27: Specify whether the 98% accuracy refers to classification of abnormal behavior overall model performance, or simulation fidelity.

The introduction is well motivated but could benefit from linking behavioral risk analysis with system-level control strategies more explicitly.

Response 1: This is a very necessary comment, and we fully agree with the reviewer's suggestion to streamline the abstract. After further discussion, we found that the abstract in the manuscript was indeed somewhat lengthy. Considering the entire research content, we have revised the abstract by removing redundant parts. Starting from the background, we introduced the necessity of the study, then presented our main research content and experimental data to demonstrate the effectiveness of this research. The 98% accuracy we mentioned refers to the accuracy of abnormal behavior classification. This result was obtained through numerous simulation experiments generating massive data, followed by multiple rounds of computation, indicating the effectiveness of our proposed method in the experimental environment. The specific problem description in the introduction indeed needed simplification, so we re-summarized the significance of this study's results to improve readability. During revision, we focused on publicly available data to highlight the research potential, directly addressing the core challenge of how to efficiently integrate high-dimensional, heterogeneous real-time perception information (such as multi-target behavior and environmental states) to build a comprehensive and fast-responding dynamic risk assessment system. Furthermore, we pointed out that the risk assessment method proposed herein not only applies to single vehicles but also meets the evaluation needs for vehicles traveling in formation. Finally, we overcame the disconnection between risk assessment and control decision-making by constructing a tightly closed loop from dynamic risk evaluation to control strategy generation. This refinement made the language more concise and highlighted the key points. [page 1, line 13-23.]

Comments 2: The literature review is extensive but predominantly focused on high-level risk models and machine learning techniques. It would benefit from additional references on behavioral performance indicators, particularly in the context of curve negotiation and driver deviation from ideal trajectories

Integration:

It may be beneficial to integrate recent contributions that focus on the definition of' synthetic driver performance indicators derived from simulated driving experiments. These indicators, often based on steering wheel behavior. longitudinal acceleration, and trajectory deviation in curves, are increasingly used to characterize deviations from standard driving models and identify critical segments in real or simulated roads.

Recent simulation-based studies have shown how performance metrics obtained in curve negotiation scenarios can reveal latent safety-critical behaviors not directly captured by classical risk indicators such as TTC or DRAC, These works often introduce multi-indicator frameworks that can complement data-driven or entropy-based models.

The development of compact, transferable performance indexes from driving simulators, which quantify drivers' adaptive responses to geometric elements (such as curves with varying radii), could serve as a valuable reference point for integrating behavioral aspects into the proposed risk assessment framework

The introduction is well motivated but could benefit from linking behavioral risk analysis with system-level control strategies more explicitly.

Response 2: We fully agree with the expert's opinion. After reviewing relevant literature, we found that in curve sections, the quantification of vehicle trajectory deviation is central to safety assessment. Lateral offset distance, as a fundamental metric, directly reflects the degree to which a vehicle deviates from the lane centerline. Some research teams, through real-vehicle tests, found that on road segments with three-dimensional curve continuity degradation, vehicles tend to systematically drift left downhill (average offset 0.35m) and significantly drift right on uphill right turns (maximum offset 1.2m). This lateral deviation is especially pronounced during turn entry. To more finely describe offset characteristics, further research on trajectory curvature variation rate has been proposed. By comparing the designed curvature with actual trajectory curvature, it reveals a phenomenon at hairpin curves where "steering is initiated earlier when entering the curve and corrected later when exiting," indicating an asymmetry in driver cornering strategy. We plan to fully incorporate these findings in future research. We are also preparing to use driving simulators to verify additional risks arising from vehicle dynamics, supplementing our focus on external and inherent vehicle risks in traffic environments. Integrating research results related to the "composite driver performance metrics derived from simulated driving experiments," as pointed out by the reviewer, is indeed beneficial. It can address some preset assumptions in our study; although many parameters require labeling in the current simulation environment, some are system defaults. In driving simulators, these parameters will be diversified through driver operation, which better supports operational risk identification. [page 4, 162-164.]

35.   Zhang C, et al. A Study on Speed Limit of Different Visibility on Expressways under Foggy Weather. Transportation Information and Safety 2023, 36(5), 25-33.

36.   Xu J, Chen Y, Zhang XB, et al. Track Curvature Characteristics and Vehicle Cornering Patterns on Hairpin Curves[J]. Journal of Southwest Jiaotong University 2021, 56(6), 1143-1152.

37.   Wang X, We X, Wang X. Investigating Micro-Driving Behavior of Combined Horizontal and Vertical Curves Using an RF Model and SHAP Analysis. Appl. Sci. 2024, 14, 2369.

Comments 3: The selection of TTC, DRAC, and MDRAC is appropriate and well-motivated. However, the paper would benefit from discussion and comparison of alternative or complementary metrics, especially those targeting driver-vehicle interaction patterns.

Integration:

The inclusion of metrics specifically designed to reflect steering stability or behavioral consistency in curve handling may enhance the depth of the proposed framework. These indicators have been used in prior studies to evaluate both infrastructure-induced risk and driver adaptation patterns under non-ideal conditions.

Furthermore, some works have proposed performance measures directly usable in ADAS or V2l contexts, highlighting their potential role in supporting proactive warnings and personalized safety alerts in smart road environments.

Response 3: This is a meaningful comment. In the research process, the commonly seen risk assessment approaches include risk metric methods, dynamic model prediction methods, machine learning methods, fuzzy logic, and expert systems. From our literature review, most of these methods are used individually and struggle to adapt to the complex urban roadway traffic environment for intelligent connected vehicle operation risk assessment. To enhance the safety and reliability of autonomous driving systems in complex environments, we proposed a comprehensive evaluation method based on multiple risk metrics, including TTC, DRAC, and MDRAC. These metrics evaluate parameters such as vehicle speed, acceleration, deceleration, time-to-collision, and headway distance. Then, weights are assigned objectively using the entropy weight method to obtain a quantified risk evaluation indicator. Our proposed approach aims to reduce the limitations of single metric methods in risk assessment by compensating for the shortcomings of one metric with others, striving for overall optimization and ensuring safer operation of autonomous driving systems. [page 8-10, line 325-400.]

Comments 4: The performance comparison with other models is valuable; still, the behavioral interpretability of risk classes could be improved by incorporating driver state references or action profiles.

Response 4: This is a very good suggestion. We have supplemented the manuscript with the methods used for the experimental datasets and the analysis of experimental results, which will help readers better understand the content. In risk quantification, we use distribution fitting tests, a statistical method to verify whether the generated sample data conforms to the theoretical distribution of traffic conflicts. We employ the Kolmogorov-Smirnov (K-S) test to compare parameters generated by the simulation system, such as minTTC, maxDRAC, and maxMDRAC, with the theoretical distribution curves, calculating the p-value to assess the goodness of fit. This quantifies the distribution characteristics of risk indicators and provides a basis for model selection. For example, when verifying if TTC data follow a normal distribution, it is necessary to calculate its shape and scale parameters and evaluate the residual sum of squares. In the presentation of specific experimental results, to demonstrate the feasibility of the study, we analyzed the position and velocity information of the target vehicle based on the vehicle trajectories and motion characteristics in the selected experimental data, then assessed the risk levels of collisions in the intersection area according to their relative positions. [page 10, line 388-400.]

Comments 5: The conclusions are well structured, though they mostly summarize the results. A more explicit discussion of limitations (e.g., generalizability, reliance on simulation, lack of real driver input) is recommended.

Suggested Future Direction:

The authors may consider as a future step extending the current model by integrating behavioral performance measures that have shown promise in identifying hidden safety-critical conditions during curve negotiation or transition zones in mixed traffic. These measures may also serve as input for driver state estimation modules or real-time feedback systems within the cloud-based architecture.

Response 5: This is a very insightful comment. This study provides a structured risk assessment and quantitative classification method for autonomous vehicles, especially for single vehicles and vehicle fleets operating in urban intersection environments, and verifies its initial effectiveness. However, there are still limitations. Future research will help advance this method toward practical application and support the development of safer and more reliable autonomous driving systems. The research approach we propose focuses on risk assessment for different levels of autonomous vehicles operating in complex urban road environments. In real-world traffic scenarios, there are various traffic participants, including human-driven cars, autonomous vehicles of different levels, non-motorized vehicles, and pedestrians. Among motor vehicles, there are trucks, motorcycles, heavy-duty vehicles, and large engineering vehicles. Road infrastructure also varies, and natural environments are constantly changing. We found that these factors are difficult to test in real traffic scenarios because they create significant experimental challenges. Therefore, we first chose to verify the method's effectiveness through simulation experiments. When building the simulation environment, constraints were placed on the road environment (including road types, lane specifications, number and width of lanes, road curvature and slope), traffic infrastructure (including traffic lights, traffic signs, road markings, cones, guardrails), traffic participants (including autonomous vehicles, human-driven vehicles, non-motorized vehicles such as bicycles and electric scooters, pedestrians, and any possible road obstacles), natural environment (including weather changes such as fog, snow, rain; lighting conditions including low light at dawn and dusk; and road surface conditions such as wetness and icing), and data and communication methods (covering Vehicle-to-Vehicle (V2V), Vehicle-to-Infrastructure (V2I), Vehicle-to-Pedestrian (V2P), and Vehicle-to-Network (V2N) communications—collectively known as V2X or Vehicle-to-Everything communication). Despite extensive foundational work, simulation environments still struggle to closely replicate real traffic conditions. Therefore, we are preparing to conduct experiments at a closed testing ground to make up for the shortcomings of our simulation experiments. [page 20, line 670-684.]

Comments 1: The primary purpose of the article is important, but the content overlaps with many similar studies in the literature. Methodological differences are not clearly highlighted.

Response 1: This is a very important comment. We compared the comprehensive risk indicator measurement method we selected in the risk level quantification experiment with other common methods, including collision time, collision avoidance deceleration rate, and improved collision avoidance deceleration rate. In the risk quantification experiment, we randomly selected 10 pieces of data to verify whether the risk level was reasonable. The experimental results showed that at a time of 3.1 seconds, three risk values of minTTC, maxDRAC, and maxMDRAC were reached, with a risk level of 3 and a high collision risk. I will also further improve the experimental data and environment, conduct experiments with larger amounts of data, so as to obtain more reliable experimental results and provide necessary data support for future real vehicle testing in actual traffic environments.

Comments 2: The original contribution should be clearly highlighted; clarify what is being proposed that is not already in the literature. The emphasis on originality should be increased, and the contributions should be clarified.

Response 2: This is a very important comment. We have clarified the main innovation of this article in the introduction section, proposing and constructing a comprehensive dynamic risk assessment and decision support framework suitable for both bicycles and formations in complex urban environments. [page 2, and 42-61]

(1) Unified multidimensional risk quantification: Overcoming the limitations of existing methods that focus on single risks or scenarios, integrating multiple dimensions of risk including collision, lateral stability, rule compliance, and traffic efficiency. It fuses multi-source information (perception, localization, maps, rules) to comprehensively characterize the complex risk states of both individual vehicles and platoons within mixed traffic flows under a unified framework.

(2) General risk assessment mechanism for single vehicles and platoons: Filling the gap in current research that mainly targets single vehicles. The proposed mechanism ap-plies not only to individual vehicles but innovatively quantifies the overall structural stability risks and inter-vehicle cooperative interaction risks in platoon mode, providing critical safety inputs for formation cooperative control.

(3) Closed-loop decision-making based on real-time risk: Addressing the deficiency of the disconnect between risk assessment and control decisions, the framework builds a tight closed loop from dynamic risk assessment to control strategy generation. This framework can generate real-time decision commands balancing safety (risk minimization) and efficiency (ensuring smooth traffic flow) according to changes in multidimensional risks (such as emergency collision avoidance, comfortable deceleration, lane change recommendations, and formation reorganization needs), significantly enhancing the rationality and informatization level of decisions in complex and dynamic environments, especially when uncertain behaviors from traffic participants exist.

Comments 3: The originality of the study is not strong enough; most of the contributions are integrations of existing studies. The methodology section should be further elaborated, and the simulation scenarios, in particular, should be clarified.

Response 3: This is a very good comment. The research method we proposed focuses on risk assessment for autonomous vehicles of different levels operating in complex urban road environments. Currently, in real traffic scenarios, there are various traffic participants, including manually driven vehicles, autonomous vehicles of different levels, non-motorized vehicles, and pedestrians. Regarding motor vehicles, there are also trucks, motorcycles, lorries, and large engineering vehicles. Road infrastructure also varies, and natural environmental conditions change as well. In our study, we found that it is very difficult to test these factors in real traffic scenarios because it would cause significant complications for the experiments. Therefore, we chose to first validate the effectiveness of the method through simulation experiments. When constructing the simulation environment, we set constraints on the road environment (including road and lane specifications, intersections, the number and width of lanes, road curvature and slope, etc.), traffic infrastructure (including traffic lights, traffic signs, road markings, cones, guardrails, etc.), traffic participants (including autonomous vehicles, human-driven vehicles, non-motorized vehicles such as bicycles and electric scooters, pedestrians, and any possible obstacles on the road), natural environment (including weather changes such as fog, snow, rain; lighting conditions such as low light at dawn and dusk; and road surface conditions such as wetness and icing), as well as data and communication methods (including vehicle-to-vehicle (V2V), vehicle-to-infrastructure (V2I), vehicle-to-pedestrian (V2P), and vehicle-to-network (V2N) communications—collectively known as V2X (Vehicle-to-Everything) communication). Despite all the foundational work done, it remains challenging for the simulation environment to closely approximate real traffic conditions, so we are preparing to conduct experiments in a closed testing ground to compensate for the limitations of our simulation experiments. [page 13-14, and line 501-527.]

Comments 4: The proposed methods do not differ methodologically from previous studies; they simply integrate the combined use of several metrics with entropy-based weighting.

Response 4: This is a meaningful comment. We analyzed the advantages and disadvantages of numerous individual risk metrics in the risk assessment of intelligent connected vehicles during driving. We then proposed the need to establish a comprehensive risk assessment method that uses models to predict the future states of vehicles and their environments, evaluate potential risks, and employ machine learning techniques to train on large amounts of simulation data to learn risk patterns and develop risk level evaluation strategies. After extensive experimental validation, we selected three effective metrics—TTC, DRAC, and MDRAC—to build a comprehensive evaluation model. The "target weighting" primarily uses the entropy weight method for objective weighting, followed by Gaussian distribution fitting in probabilistic risk assessment for risk quantification and risk level classification. The quantification methods for key evaluation indicators in the three risk assessment approaches are provided, dividing the risk levels into three grades—"1, 2, 3"—based on vehicle driving conditions in urban road environments. Through the theoretical research described above, risk identification and risk level evaluation were completed. Then, the method's effectiveness was further validated through simulation environments, providing detailed data support for future testing in real traffic environments.

[page 9-10, line 363-400]

Comments 5: Comparisons should be made to demonstrate the superiority of entropy-based weighting with alternative methods.

Response 5: This is a very important comment. When evaluating the dynamic risk of intelligent connected vehicles during operation in complex urban road environments and quantifying it, it involves dynamic characteristics such as vehicle motion behavior and driving trajectories, as well as static road traffic conditions and other road users. Many researchers also choose methods like Analytic Hierarchy Process (AHP), Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS), or entropy-based algorithms to quantify risk. We have also chosen the entropy weight method for an objective evaluation of multidimensional comprehensive risk, which avoids the drawback of previous studies that applied the entropy weight method to assess single risk indicators. [page 9, line 343-362.]

Comments 6: The use of indicators such as entropy-based weighting, TTC, and DRAC are common in the literature. The authors reused them, but it is not clear that a unique framework has been presented.

Response 6: This is a great comment. The literature review in the manuscript on the use of reinforcement learning or Bayesian inference for real-time integration of multiple risk factors is indeed quite old, so we have updated some of the references. Scholars have proposed a collision avoidance method for autonomous driving vehicle clusters based on hybrid reinforcement learning in complex and changing traffic environments. This method combines the adaptive ability of reinforcement learning with the feature extraction ability of deep learning, which can extract features from environmental perception data and predict possible collision risks to improve collision avoidance performance in complex traffic scenarios. In response to the insufficient research on lane risk at urban signal controlled intersections and the uncertainty caused by complex interactions, a comprehensive evaluation framework for lane risk based on Bayesian inference and XGBoost has been established. The established lane risk comprehensive evaluation model performs better than the benchmark model in identifying medium and high risks when evaluating three types of interaction conflicts: non motorized vehicle vehicle, pedestrian vehicle, and pedestrian non motorized vehicle, especially in determining extremely dangerous interactions, which is more reasonable and avoids the overestimation problem of the benchmark model. These research directions have been considered in our manuscript and will be further reflected in our later practical traffic environment verification environment.

[page 10, line 388-400.]

Comments 7: Current sources should be added to the literature review. Some repetitions and excessive details should be simplified in the literature summary.

Response 7: I strongly agree with the comment. After our discussion, we found that there was indeed content duplication. We revised the introduction section to clarify our research objectives, research methods, and key work. We also modified the literature review section to summarize the current situation, highlighting the shortcomings of existing research results and laying a clear foundation for further detailed research in the following sections.

Comments 8: The experimental design and scenario descriptions provided in the CARLA-SUMO simulation platform remain superficial.

Response 8: This is a very important comment. Indeed, we found in the research process that the CARLA-SUMO environment was a very good tool to realize the simulation of autonomous vehicle running in the urban road environment. We spent a lot of work in the experiment process to design the road environment, including the specifications of intersections, the number and width of lanes, the curvature and gradient of roads, etc. These parameters need to be precise to practical applications in order to reproduce real-world driving conditions in simulations. There is also transportation infrastructure, including all physical and non physical facilities used to support smooth and safe transportation. This includes but is not limited to traffic lights, traffic signs, road markings, cones, guardrails, etc. These infrastructures not only need to be designed according to actual specifications and layouts, but also need to simulate their functions in different situations, such as the changing patterns of traffic lights, the visibility of traffic signs, and the difficulty of recognition under different lighting and weather conditions. These factors will all have an impact on the decision-making system of autonomous vehicles. At the same time, the design of traffic participants refers to all entities moving on the road, including autonomous vehicles, human driven vehicles, non motorized vehicles (such as bicycles, electric scooters), pedestrians, and any obstacles that may appear on the road. Because there are many parameters involved in the debugging process, we did not list them in the manuscript. If readers need them, they can contact the corresponding author for assistance. We have added a schematic diagram of the joint simulation environment in the article. [page 13-14, line 501-527.]

Comments 9: A 98% success rate is claimed; however, how the results were evaluated and the validation method are inadequately explained. Clearly state which dataset and scenario yielded the accuracy rate.

Response 9: This is a very necessary comment, and we fully agree with the reviewer's suggestion to streamline the abstract. After further discussion, we found that the abstract in the manuscript was indeed somewhat lengthy. Considering the entire research content, we have revised the abstract by removing redundant parts. Starting from the background, we introduced the necessity of the study, then presented our main research content and experimental data to demonstrate the effectiveness of this research. The 98% accuracy we mentioned refers to the accuracy of abnormal behavior classification. This result was obtained through numerous simulation experiments generating massive data, followed by multiple rounds of computation, indicating the effectiveness of our proposed method in the experimental environment. The specific problem description in the introduction indeed needed simplification, so we re-summarized the significance of this study's results to improve readability. During revision, we focused on publicly available data to highlight the research potential, directly addressing the core challenge of how to efficiently integrate high-dimensional, heterogeneous real-time perception information (such as multi-target behavior and environmental states) to build a comprehensive and fast-responding dynamic risk assessment system. Furthermore, we pointed out that the risk assessment method proposed herein not only applies to single vehicles but also meets the evaluation needs for vehicles traveling in formation. Finally, we overcame the disconnection between risk assessment and control decision-making by constructing a tightly closed loop from dynamic risk evaluation to control strategy generation. This refinement made the language more concise and highlighted the key points. [page number, paragraph, and line.]

[page number, paragraph, and line.]

Comments 10: The presentation and interpretation of analysis outputs are weak.

Response 10: This is a very good suggestion. We should enrich the interpretation of the experimental results, fully demonstrate the excellent implementation results and data performance, and make it easier for readers to understand what our research results mainly include, how to achieve risk judgment through vehicle motion status, and provide guarantees for further command of cloud control system strategy generation. We have revised the interpretation of the first experimental result. [page 18, line 606-615] We have revised the interpretation of the second experimental result. [page 19, line 629-638]We have revised the interpretation of the third experimental result. [page 20, line 654-663]

Comments 11: Fuzzy logic, expert systems, and ML methods are briefly mentioned; however, how and when these were chosen is unclear within the methodological framework.

Response 11: This is a meaningful comment. Common risk assessment methods encountered during our research include risk metric indicator methods, dynamic model prediction methods, machine learning methods, fuzzy logic, and expert systems. From our review of relevant literature, these methods are mostly used individually, making it difficult to adapt to risk assessment for intelligent connected vehicles driving in the complex environment of urban road traffic. To improve the safety and reliability of autonomous driving systems in complex environments, we propose a comprehensive assessment method based on multiple risk metric indicators, including TTC, DRAC, and MDRAC. [page 4, line 190-203] These are used to evaluate parameters such as vehicle speed, acceleration, deceleration, time headway, and headway distance. Then, weights are assigned objectively using the entropy weight method to obtain a quantified risk assessment indicator. The method we propose aims to reduce the shortcomings of the original single risk metric indicator method in risk assessment by compensating for the weaknesses of one indicator with others, striving to achieve overall optimization and ensure safer operation of autonomous driving systems. [page 15, line 554-561)

Comments 12: The references are generally appropriate, but some sources are very old, and current AI-based risk assessment approaches are lacking.

Response 12: Totally agree. During our research, there were several references that were indeed quite old. After the reviewers' suggestions, we reorganized the domestic and international research results and replaced the relatively old papers with newer ones to make up for the shortcomings of our previous work.

3.     Pang et al. Risk Assessment Method for Autonomous Vehicles Violating Safety Common Sense Based on Driving Behavior. IEEE Access 2025, 13, 63076-63092.

10.   Fei M, Xu W, Wei HY. Real-time accident risk identification for freeway weaving segments based on video analytics. Meas-urement 2025, 242, 115783.

18.   Pawar NM, Arkatkar S. Modeling drivers' dilemma at unsignalized T-intersections under mixed traffic conditions: A case study from India. Traffic Injury Prevention 2025, 25(8), 1107-1114.

At present, new methods such as artificial intelligence technology and large language models are widely used in the academic community. We also used similar methods in our research process, but it was not in-depth enough. We need to further study more excellent results. We have supplemented representative literature in the article.

38.   Hu CF, Li XD, Pang JX. Trustworthy Driver State Perception via Contextual Interaction-Driven Evidential Vision-Language Fusion in Vehicular Cyber-Physical Systems. IEEE T-ITS 2025, 26(2), 2202-2211.

39.   Qing CL, Ruo HY, Ying FC. Collision risk prediction and takeover requirements assessment based on radar-video integrated sensors data: A system framework based on LLM. Accident Analysis & Prevention 2025, 218, 108041.

Comments 13: The conclusion section should be rewritten. The conclusions should be supported by more robust comparisons and statistical analyses.

Response 13: This is a very insightful comment. This study provides a structured risk assessment and quantitative classification method for autonomous vehicles, especially for single vehicles and vehicle fleets operating in urban intersection environments, and verifies its initial effectiveness. However, there are still limitations. Future research will help advance this method toward practical application and support the development of safer and more reliable autonomous driving systems. The research approach we propose focuses on risk assessment for different levels of autonomous vehicles operating in complex urban road environments. In real-world traffic scenarios, there are various traffic participants, including human-driven cars, autonomous vehicles of different levels, non-motorized vehicles, and pedestrians. Among motor vehicles, there are trucks, motorcycles, heavy-duty vehicles, and large engineering vehicles. Road infrastructure also varies, and natural environments are constantly changing. We found that these factors are difficult to test in real traffic scenarios because they create significant experimental challenges. Therefore, we first chose to verify the method's effectiveness through simulation experiments. When building the simulation environment, constraints were placed on the road environment (including road types, lane specifications, number and width of lanes, road curvature and slope), traffic infrastructure (including traffic lights, traffic signs, road markings, cones, guardrails), traffic participants (including autonomous vehicles, human-driven vehicles, non-motorized vehicles such as bicycles and electric scooters, pedestrians, and any possible road obstacles), natural environment (including weather changes such as fog, snow, rain; lighting conditions including low light at dawn and dusk; and road surface conditions such as wetness and icing), and data and communication methods (covering Vehicle-to-Vehicle (V2V), Vehicle-to-Infrastructure (V2I), Vehicle-to-Pedestrian (V2P), and Vehicle-to-Network (V2N) communications—collectively known as V2X or Vehicle-to-Everything communication). Despite extensive foundational work, simulation environments still struggle to closely replicate real traffic conditions. Therefore, we are preparing to conduct experiments at a closed testing ground to make up for the shortcomings of our simulation experiments.

[page 20, line 670-684.]

Comments 14: Risk indicators are detailed, but the reason why entropy-based weighting is the most appropriate method is not discussed.

Response 14: This is a meaningful comment. We analyzed the advantages and disadvantages of numerous individual risk metrics in the risk assessment of intelligent connected vehicles during driving. We then proposed the need to establish a comprehensive risk assessment method that uses models to predict the future states of vehicles and their environments, evaluate potential risks, and employ machine learning techniques to train on large amounts of simulation data to learn risk patterns and develop risk level evaluation strategies. After extensive experimental validation, we selected three effective metrics—TTC, DRAC, and MDRAC—to build a comprehensive evaluation model. The "target weighting" primarily uses the entropy weight method for objective weighting, followed by Gaussian distribution fitting in probabilistic risk assessment for risk quantification and risk level classification. The quantification methods for key evaluation indicators in the three risk assessment approaches are provided, dividing the risk levels into three grades—"1, 2, 3"—based on vehicle driving conditions in urban road environments. Through the theoretical research described above, risk identification and risk level evaluation were completed. Then, the method's effectiveness was further validated through simulation environments, providing detailed data support for future testing in real traffic environments.

[page 9-10, line 343-400.]

Comments 15: Figure 2 should be redesigned to increase its visual appeal.

Response 15: This is a very good suggestion, and it is indeed necessary to further optimize the design of Figure 2 in the manuscript. After further modifications, it is easier to read and understand, and the execution effect of the operation control strategy can be more clearly defined. We have divided three road scenarios based on the driving environment of vehicles, namely the entrance section of the intersection, the entrance lane of the intersection, and the internal space of the intersection. Vehicles exhibit different behavioral characteristics in different scenarios, such as the entrance section of the intersection, which is usually in a free flow state. Vehicles can follow, change lanes, overtake, and other behaviors according to expectations; At the entrance of the intersection, the traffic density increases, the coupling relationship between vehicles strengthens, the frequency of vehicle acceleration and deceleration increases, and the typical behavioral characteristics of vehicles are following the vehicle while maintaining a safe distance between them; The internal space of an intersection, where vehicles are in a state of passing through the intersection, is influenced by the driving route (left turn, straight turn, right turn), geometric topology of the intersection, and conflicts between vehicles in other directions of entry. Therefore, the adjusted diagram can better express the selection of reasonable, safe, and feasible cloud takeover strategies for vehicles based on their different scenarios and states. [page 12, line 444.]

Comments 1: As a final step before publication, I kindly suggest that the authors format the manuscript according to the journal's official template. This includes aligning the layout, headings, references, figures, tables, and any supplementary materials with the journal's guidelines.

Response 1: We sincerely thank you for your detailed suggestions on the content and formatting of the manuscript! We have carefully revised it and strictly followed the formatting requirements of this journal to complete the standardization process.

Comments 1: Introduction section in its current form is not an appropriate article format. It is recommended that it be revised to conform to the appropriate article format.

Response 1: This is a very important opinion. The introduction section is the key beginning of an academic paper. We further revised the introduction part of the paper based on the comments of the reviewers. Through the digital display in California, we introduced the background of our research, further clarified the research issue of the lack of a risk evaluation index system for autonomous vehicles, and clearly defined the main content of our research and the future value contribution. Thus, readers can quickly understand the research focus of our paper. [page 1-2, line 27-45]

Comments 2: Similarly, the conclusion section should be rewritten.

Response 2: We are very grateful for the comments from the reviewers. We carefully discussed the contents that should be included in the conclusion writing. Our main contribution was proposing a comprehensive risk assessment framework for autonomous driving on urban roads, integrating multiple traditional evaluation indicators, and introducing a cloud-assisted takeover strategy for the safe operation of high-level autonomous vehicles. Then, we used simulation to verify the effectiveness of the existing design strategies. Of course, we also found that our research still has room for improvement. We should focus on high-fidelity complex scenario verification and improve the risk mitigation strategies under unstable communication to provide a foundation for real environment testing. [page 20, line 653-662]

Comments 3: The entire article, from beginning to end, should be formally rewritten according to scientific article writing guidelines. Formatting errors and grammatical errors should be corrected accordingly.

Response 3: Thank you very much. This is a very important comment. We have re-employed a professional English editing expert to help us correct the formatting and grammar issues. They did point out some problems, and we have carefully revised them. This has been of great help for the subsequent writing of our paper.

Comments 4: Figure 4 should be prepared in readable fonts.

Response 4: Yes, we have changed the font of the text in Figure 4 to "bold", which makes it easier to read. At the same time, we have re-adjusted the resolution of the images, making the content easier to read. The size of the axis labels was also further adjusted. [page 16, line 546]

Comments 5: The abbreviations used in the article should be used appropriately according to the rules. Using abbreviations in subheadings is generally discouraged (e.g., TTC, DRAC, etc.).

Response 5: This is a very important suggestion. We have revised all the expressions in the manuscript as per the requirements, standardized the abbreviations of terms in the titles, and also corrected the other abbreviations in the manuscript to avoid repetition and redundancy. [page 5-6]

Comments 6: Paragraph spacing should be formatted appropriately (e.g., 291-292).

Response 6: Thank you very much to the reviewer. This is a very important comment. We have re-modified all the paragraph spacing according to the requirements of the journal, strictly checked and revised all the line spacing, making our manuscript more standard.

Comments 7: Line spacing in Table 1 should be wide enough.

Response 7: This is a very good suggestion. The line spacing of Table 1 in our manuscript is indeed too narrow and does not follow the format specified by the journal properly. We have re-reviewed the formatting in other parts and corrected all the non-compliant areas.

Comments 8: Since abbreviations such as vehicle-to-vehicle (V2V) are used only once, there is no need for abbreviations.

Response 8: This is a very good suggestion. The abbreviation of professional terms should only appear once. We re-reviewed the formatting of the entire manuscript and found that there were several repeated abbreviations. We have re-optimized and revised them.