Intelligent Vehicle Driving Risk Assessment in Multi-Vehicle Interaction Scenarios: A Comprehensive Review

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Dec 15 2023

In a paper published in the journal World Electric Vehicle, researchers delved into automotive intelligence, acknowledging significant strides in the field while grappling with persistent skepticism, particularly regarding intelligent vehicles navigating complex multi-vehicle environments.

The paper synthesized existing literature on data collection methods, vehicle interaction dynamics, and risk evaluation techniques in these scenarios. By critiquing limitations and envisioning future developments, the study offered crucial guidance for enhancing safety measures in real-world multi-vehicle settings for intelligent vehicles.

Background

The surge in automotive intelligence spurred by AI and manufacturing advancements has elevated research focus within the automotive industry. Intelligent vehicles, leveraging cutting-edge sensors and technologies, face skepticism, especially in intricate multi-vehicle environments with heightened uncertainties and driving risks. Existing risk evaluation metrics like Time-to-X indicators often need to catch up in capturing the complexity of these scenarios, leaving a void in comprehensive assessments for intelligent vehicles navigating multi-vehicle interactions.

Data Collection in Multi-Vehicle Scenarios

The initial step for conducting risk assessments in multi-vehicle interaction scenarios is to gather offline interaction data among multiple vehicles to develop and train risk evaluation models. Subsequently, researchers fed real-time interaction data into these models for live risk assessments. Data collection methods in this domain primarily fall into three categories: vehicle-based, roadside-based, and simulation-based approaches.

Recent advancements in onboard software/hardware and intelligent algorithms have expanded access to autonomous driving datasets, eliminating constraints imposed by experimental environments and technical limitations. These datasets encompass various aspects: vehicle-related data detailing operational, driver behavior, and internal vehicle status; driving environment data offering comprehensive perspectives on traffic participant states, high-definition mapping, and behavior classification; and complex interaction behavior datasets focusing on intricate traffic scenarios to support decision-making development. These diverse datasets serve as invaluable resources for conducting multi-vehicle interaction risk analyses.

Multi-Vehicle Interaction Behavior Analysis

Understanding the interaction behavior among multiple vehicles is crucial for intelligent cars to navigate complex environments and accurately assess risks. Previous research has extensively analyzed behavior within vehicle pairs, yet the broader context involves the influence of entire vehicle groups, resulting in more intricate and diverse interaction patterns.

Delineating typical interaction behavior categories—longitudinal, front cut-in, rear cut-in, and lateral behaviors—observed in vehicle pairs, the scope needs expansion to encompass complexities arising from interactions among all surrounding vehicles. This comprehensive analysis is crucial for a deeper understanding of multi-vehicle interaction dynamics.

Risk Assessment Overview

A comprehensive review of risk assessment methods pertinent to multi-vehicle interaction scenarios starts by delineating traditional approaches for assessing risks within two-vehicle interactions. Subsequently, it explores existing methods tailored for evaluating risks within multi-vehicle interactions.

Interaction between Two Vehicles: Quantifying the risk in two-vehicle interactions involves parameters like vehicle proximity and collision avoidance intensity, establishing Surrogate Safety Measures (SSM). These SSMs encompass time-based, distance-based, deceleration-based, and energy-based categories. Time-based SSMs, including Time to Collision (TTC) and Post Encroachment Time (PET), gauge proximity to potential and actual collision points but need more depth in describing the collision process. Distance-based SSMs measure risk by distances to collision points.

At the same time, deceleration-based and energy-based SSMs evaluate evasive actions and collision severity based on braking capabilities and energy released during potential collisions. These methods, though efficient, often need to pay more attention to uncertainties in vehicle motion and interaction scenarios, constraining their applicability in complex multi-vehicle contexts.

Multi-Vehicle Interaction: This segment categorizes existing multi-vehicle interaction risk assessment methods into state inference-based and trajectory prediction-based methods. State inference methods utilize historical trajectories for direct risk depiction without predicting future trajectories, employing theories like causation, interactive decision-making, and distribution relationships. Causal methods, using Bayesian models, dynamically input kinematic parameters for risk prediction. Interactive decision-making methods, leveraging game theory, analyze subject-adjacent vehicle pairs but may overlook broader gaming behaviors. Distribution relationship methods employ complex network theory but currently lack empirical data.

Evaluation Method Based on Trajectory Prediction: Trajectory prediction-based methods forecast vehicle trajectories and assess multi-vehicle interaction risks based on these predictions. Field theory-based evaluations employ artificial potential fields or probabilistic driving risk fields to quantify interaction risks. Spatiotemporal proximity assessments, relying on indicators like Time to Collision (TTC), quickly evaluate risks but may neglect uncertainties in surrounding vehicles. Methods addressing uncertainty encompass interactions, model parameters, and predicted trajectories, refining risk assessment accuracy but often demanding higher computational resources and precise behavioral modeling.

Overall Assessment of Trajectory Prediction-based Methods: Trajectory prediction-based assessments, while computationally intensive, offer more accurate evaluations due to comprehensive consideration of future spatial relationships among interacting vehicles. By addressing a broader scope of uncertainties, these methods tend to have longer prediction horizons and can identify collision risks earlier.

Conclusion

To sum up, addressing the complex task of safety assessment within multi-vehicle interaction scenarios, this article systematically reviewed the literature concerning data collection methodologies, vehicle interaction mechanisms, and prevailing multi-vehicle interaction risk assessment methods. The overview categorized these methodologies into state inference-based and trajectory prediction-based methods, highlighting their advantages, limitations, and applicability within multi-vehicle settings.

Moreover, discussions emphasized the challenges within this domain, focusing on incomplete environmental information, method robustness, validation complexities, scope determination, and the need to broaden factors influencing risk assessment accuracy. Further exploration of methodologies like clustering knowledge graphs could offer a more comprehensive understanding of research focal points and developmental trajectories within risk assessment in multi-vehicle interaction scenarios.