Learning collision risk proactively from naturalistic driving data at scale

TL;DR

This paper introduces Generalised Surrogate Safety Measure (GSSM), a data-driven framework that proactively quantifies collision risk from naturalistic driving data without relying on crash labels. By modelling context-conditioned spacings with a lognormal distribution and learning conditional parameters through neural encoders, GSSM provides a continuous risk score and allows early, reliable alerts across diverse urban and highway interactions. The approach is demonstrated on a large SHRP2-based dataset, augmented with ArgoverseHV and highD data, achieving high AUPRC (~0.90) and competitive, stable alert timeliness (median TTI around 2.5 s). GSSM’s context-aware design, scalability, and attribution capability (via Expected Gradients) support proactive safety in autonomous systems and traffic management, with open access to code and data for reproducibility and further development.

Abstract

Accurately and proactively alerting drivers or automated systems to emerging collisions is crucial for road safety, particularly in highly interactive and complex urban environments. Existing methods either require labour-intensive annotation of sparse risk, struggle to consider varying contextual factors, or are tailored to limited scenarios. Here we present the Generalised Surrogate Safety Measure (GSSM), a data-driven approach that learns collision risk from naturalistic driving without the need for crash or risk labels. Trained over multiple datasets and evaluated on 2,591 real-world crashes and near-crashes, a basic GSSM using only instantaneous motion kinematics achieves an area under the precision-recall curve of 0.9, and secures a median time advance of 2.6 seconds to prevent potential collisions. Incorporating additional interaction patterns and contextual factors provides further performance gains. Across interaction scenarios such as rear-end, merging, and turning, GSSM consistently outperforms existing baselines in accuracy and timeliness. These results establish GSSM as a scalable, context-aware, and generalisable foundation to identify risky interactions before they become unavoidable, supporting proactive safety in autonomous driving systems and traffic incident management. Code and experiment data are openly accessible at https://github.com/Yiru-Jiao/GSSM.

Figures (10)

• Figure 1: Statistics of traffic accidents. (a) Estimated numbers of fatalities due to road traffic accidents per year by the Institute for Health Metrics and Evaluation GBD2022. (b) Distribution of traffic accidents that occurred in different locations from 2001 to 2021 in 27 countries where complete data are accessible. The data is sourced from the United Nations Economic Commission for Europe (UNECE) Statistical Database dataUNECE.
• Figure 2: The safety pyramid conceptualises the evolution from safe interactions to unsafe interactions up to crashes.
• Figure 3: Statistics of original and processed events in the SHRP2 NDS. (a) Numbers of crashes, near-crashes, and safe baselines that are recorded, reconstructed, and used in the test. (b) Distribution of event types that are recorded, reconstructed, and used in the test. (c) Distribution of weather and road surface conditions in the test events. (d) Distribution of lighting conditions in the test events. (e) Distribution of traffic conditions in the test events. LOS stands for level of service, details of which are referred to Appendix Table \ref{['tab: context features']}.
• Figure 4: Separation of time periods for each safety-critical event in the test set for evaluation.
• Figure 5: Performance curves comparing GSSM and existing two-dimensional surrogate safety measures (2D SSMs). The GSSM under comparison is h-C which is trained on the highD data and uses current features. For the other 2D SSMs, TAdv is abbreviated for Time Advantage; ACT for Anticipated Collision Time; TTC2D for Two-dimensional Time-to-Collision; and EI for Emergency Index. Each method is tested on the same 2,591 safety-critical events, including 2,481 near-crashes and 110 crashes. In the ATC plot, the shaded bands represent 99% confidence intervals for median time to impact.