Risk Assessment of Roundabout Scenarios in Virtual Testing Based on an Improved Driving Safety Field

Wentao Chen, Aoxue Li, Haobin Jiang2024 · DOI

TL;DR

This paper introduces an improved "driving safety field" model to better assess risks in complex roundabout scenarios for autonomous vehicles. By incorporating dynamic traffic interactions and using a DENCLUE-like clustering approach, the method significantly enhances the detection of dangerous situations compared to traditional metrics like Time to Collision (TTC).

What Problem Does This Paper Address?

Autonomous vehicles (AVs) are rapidly advancing, but ensuring their safety, especially in complex traffic environments like roundabouts, remains a significant challenge. Traditional methods for assessing driving risk, such as Time to Collision (TTC), often fall short in accurately quantifying dangers in these dynamic, multi-vehicle scenarios. Roundabouts, with their unique traffic flow, merging, and diverging patterns, present a particular hurdle. Existing risk assessment models struggle to capture the intricate interactions between vehicles, leading to an inability to precisely identify and generate truly dangerous scenarios for virtual testing. This gap hinders the thorough validation of AV safety systems.

Methods

The core of this paper is an improved driving safety field model. This model builds upon the concept of a "field" in physics, where risk is distributed spatially and temporally. It comprises two main components:

  • Road Safety Field: This component accounts for the restraining effect of lane markings, adapted to the specific context of roundabouts (lane centerlines and road boundaries).
  • Interactive Safety Field: This is the novel part, designed to capture the impact of both static and dynamic traffic participants. It draws inspiration from electric field theory and incorporates factors like vehicle speed, heading angle, and the "pseudo-distance" to amplify longitudinal risks. The model also accounts for weather conditions and vehicle type.

To calibrate the parameters of this improved safety field, the authors utilize the social force model, which describes interactions between agents. They then employ a DENCLUE-like clustering algorithm to process the safety field data. This algorithm identifies "collision systems" (clusters of interacting vehicles) and calculates their risk degree. The risk is assessed both temporally (overall scenario risk) and spatially (risk within a collision system), providing a more comprehensive evaluation.

Datasets

The study utilizes the rounD dataset. This dataset contains public traffic data (position, speed, acceleration, and object type) captured from a bird's-eye view at roundabouts. It includes over 6 hours of video and data from more than 13,000 road users. The authors specifically focused on vehicle trajectories within the roundabout and processed the data to repair truncated trajectories and resample at 25 Hz.

Key Insights and Ideas

  • Roundabouts are uniquely challenging: The paper highlights that the complex merging and diverging dynamics in roundabouts are not well-captured by traditional, simpler risk metrics.
  • Field theory for risk: Applying field theory concepts to model driving risk provides a more holistic view, considering the spatial and temporal distribution of potential hazards.
  • DENCLUE for collision systems: Using a density-based clustering approach like DENCLUE is effective for grouping interacting vehicles into "collision systems," allowing for a more nuanced assessment of multi-vehicle risks.
  • Temporal vs. Spatial Risk: Differentiating between the overall temporal risk of a scenario and the spatial risk within specific interacting vehicle groups offers a powerful way to filter and prioritize critical scenarios for testing.

Key Results

The proposed method demonstrated a significant improvement in detecting dangerous scenarios. Compared to the traditional Time to Collision (TTC) method, the improved driving safety field model detected approximately 175% more dangerous scenarios. The paper also shows that the model's predicted vehicle parameters (displacement, speed, acceleration) closely match actual values, with low RMSEs (e.g., around 0.55-0.63 for displacement). In the simulated case studies, the proposed risk degree consistently identified risks during merging and multi-vehicle conflicts, while TTC often failed to exceed its threshold.

Limitations and Future Work

The authors acknowledge that the model's transferability to diverse traffic environments has not yet been fully established. The current research is focused specifically on roundabout scenarios. Future work will involve adapting the model for other complex traffic environments, such as signal-controlled intersections, and will require accounting for additional constraints like traffic lights using a force-based approach. Further refinement to accommodate the unique characteristics of different scenarios is also planned.

Summary updated

BibTeX
@misc{chen2024risk,
  author = {Wentao Chen and Aoxue Li and Haobin Jiang},
  title = {{Risk Assessment of Roundabout Scenarios in Virtual Testing Based on an Improved Driving Safety Field}},
  year = {2024},
  doi = {10.3390/s24175539},
}

Loading related content...

Figure