Driving risk emerges from the required two-dimensional joint evasive acceleration
DOI: 10.48550/arxiv.2604.17841
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the fundamental limitations of Time-to-Collision (TTC), the dominant paradigm for quantifying driving risk in autonomous systems. The authors argue that TTC is flawed because it treats risk as a one-dimensional closing problem, failing to capture the inherently two-dimensional nature of collision avoidance. This "dimensionality mismatch" leads to limited risk informativeness, where scenes with vastly different evasive difficulties receive identical TTC values, and risk-time misalignment, where risk metrics may peak just as a conflict resolves. To resolve these issues, the authors propose Evasive Acceleration (EA), a hyperparameter-free, physically interpretable metric that quantifies risk as the minimum magnitude of a constant relative acceleration vector required to alter relative motion and ensure a collision-free outcome across all possible directions. The study validates EA using a large, heterogeneous dataset comprising 44,180 naturalistic interactions from five open-source datasets (covering highways and urban intersections in Germany, China, and the US) and 658 reconstructed real-world crashes from the CIMSS-TA database. The authors conducted three complementary experiments to assess statistical separability, early-warning timeliness, and information retention. For statistical separability, they compared EA against baselines including TTC, TTC2D, ACT, DRAC, and MEI across four precrash lead windows. For warning timeliness, they calibrated thresholds based on the 90th, 95th, 99th, and 99.5th percentiles of noncrash event distributions to ensure equal false-alarm constraints. Finally, they measured information retention using a cross-entropy framework to determine how much uncertainty about crash outcomes each method reduced. The results demonstrate that EA significantly outperforms existing methods. In statistical separability tests, EA achieved the highest Area Under the Precision-Recall Curve (AUPRC) and Area Under the Receiver Operating Characteristic Curve (AUROC) across all precrash windows. For instance, in the [−1.5, −0.1] s window, EA attained an AUPRC of 0.901 and AUROC of 0.934, substantially higher than the best baselines. Regarding early warnings, EA provided the earliest statistically significant warnings under all percentile-aligned thresholds. At the strict 99.5th percentile threshold, EA warned 120–267% earlier than TTC-based methods. Furthermore, EA improved information retention by 54.2–241.4% over baselines. The study also found strong directional asymmetry in incremental information: adding EA to existing methods yielded 17.5–95.5 times more information gain than adding existing methods to EA, indicating that EA captures nearly all outcome-relevant information present in current metrics while contributing substantial nonredundant data. The significance of this work lies in establishing EA as a more faithful quantification of multidirectional interaction risk. By capturing the continuous evolution of risk and the minimum intervention cost, EA provides a physically intuitive foundation for next-generation autonomous driving systems. The findings suggest that shifting from one-dimensional TTC paradigms to two-dimensional EA can improve safety benchmarks, reduce false alarms, and enhance the training and evaluation of autonomous vehicles by providing a more accurate gradient of collision risk.
Key finding
Evasive acceleration provides earlier warnings and superior statistical discrimination of crash outcomes compared to time-to-collision and other existing risk metrics by capturing the two-dimensional nature of collision avoidance.
Methodology
dataset
Sample size: 44838
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 1 | 2026-06-03 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-03 |
| chunk | success | chunk | — | — | 1 | 2026-06-03 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-03 |
| enrich | skipped | — | — | — | 2 | 2026-06-03 |
| promote | success | — | — | — | 1 | 2026-06-03 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
- telematics crash prediction
- induced exposure
- time to collision looming
- naturalistic crash near crash
- perception reaction time
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Empirical Findings: crash risk outcomes
- Methodological Resource: dataset resource
- Theoretical Contribution: computational model