TL;DR
This paper introduces a novel approach to predicting highway crash risks by analyzing vehicle groups (VGs) and their interactions, moving beyond traditional segment-level analysis. By employing an impact-based grouping method and a Graph Neural Network (GNN) with explainability features, the study achieves high prediction accuracy and provides insights into the critical factors driving crash risk.
What Problem Does This Paper Address?
Predicting highway crash risks has traditionally relied on aggregate traffic parameters and road segment characteristics. However, this approach often overlooks the dynamic interactions between individual vehicles and how these interactions evolve over time, which are crucial for understanding and preventing crashes. The advent of advanced sensing technologies like UAVs and V2X communication allows for high-resolution trajectory data, opening up opportunities for more granular, vehicle-centric risk analysis. This paper addresses the challenge of leveraging this rich data to predict crash risks at a vehicle-group level and to interpret the underlying mechanisms driving these risks.
Methods
The study proposes an impact-based vehicle grouping (IVG) method to cluster vehicles into VGs. This method identifies vehicles that react to each other's erratic behaviors, such as sudden braking or lane changes, by evaluating their response deceleration rates. VGs are then associated across consecutive timestamps to track their evolution.
The risk within a VG is quantified using inverse Time-to-Collision (iTTC), a surrogate safety measure that captures the potential for collision between vehicle pairs. Two primary prediction models are developed:
- Logistic Regression (LR): This model uses aggregated VG information (e.g., average speed, percentage of large vehicles, current risk indicator) to predict two binary outcomes: Vehicle Group Risk Level (VGRL - high-risk or not) and Trend of Risk Scale Change (TRSC - escalated or not).
- Graph Neural Network (GNN): This model utilizes disaggregated VG information, representing each vehicle as a node and their interactions as edges in a graph. An Edge-Conditioned Convolutional (ECC) GNN is employed to capture the complex relationships and edge attributes (iTTC).
To interpret the GNN's predictions, an explainable framework is introduced, combining GNNExplainer to identify critical subgraphs and node features, and feature perturbation to assess the directional impact of these features on VG risk predictions.
Datasets
The study utilizes the MAGIC Dataset, which comprises approximately 3 hours of aerial video data collected by Unmanned Aerial Vehicles (UAVs) at six urban highway locations. This dataset provides detailed vehicle trajectories, including ID, type, position, speed, acceleration, timestamp, lane ID, and direction, captured at a frequency of 25 Hz. For consistency, only eastbound traffic data was used.
Key Insights and Ideas
- Vehicle Groups as the Unit of Analysis: Shifting the focus from road segments to vehicle groups allows for a more nuanced understanding of crash risks driven by dynamic interactions.
- Impact-Based Grouping: The IVG method effectively captures how vehicles respond to each other's behaviors, forming groups that are relevant for safety analysis.
- GNNs for Disaggregated Data: GNNs are well-suited to model the complex, relational structure of vehicle groups, outperforming traditional methods when using disaggregated, individual vehicle data.
- Explainability is Crucial: The combination of GNNExplainer and feature perturbation provides valuable insights into why a GNN predicts a certain risk level, identifying critical features and their directional influence.
- Velocity is Key: Vehicle velocity is identified as the most critical feature for predicting VGRL and TRSC, suggesting that speed dynamics play a significant role in crash risk.
- VGs Tend to Persist: The "risk_indicator" feature in the LR model shows that VGs already at high risk are likely to remain so, highlighting the importance of early intervention.
Key Results
Both the LR and GNN models achieved excellent prediction performance, with AUC values exceeding 0.93 for VGRL and 0.86 for TRSC at a 1-second prediction interval. The performance generally improved as the prediction time interval decreased.
The GNN model, using disaggregated data, achieved AUCs of approximately 0.93 for VGRL and 0.85 for TRSC at a 1-second prediction interval.
The explainability framework revealed that:
- Vehicle velocity (v) has a negative impact on VG risk predictions, while other features like acceleration (abs_a), type, lane change (changing_lane), and proximity to on/off-ramps tend to have a positive impact.
- For VGRL prediction, significant node features are concentrated at the head of VGs, indicating that leading vehicle behaviors are critical.
- For TRSC prediction, significant features are more distributed towards the center of VGs, suggesting that interactions within the group can escalate risk.
Limitations and Future Work
The authors acknowledge that the study's risk indicators could be enriched by integrating additional factors such as driver habits, weather conditions, and road surface conditions. This would further enhance the analysis and yield more robust insights. Future work could explore these additional factors to develop more comprehensive and accurate crash prediction models.
Summary updated