SECURE: Stable Early Collision Understanding via Robust Embeddings in Autonomous Driving

Wenjing Wang; Wenxuan Wang; Songning Lai

cs.LGApr 1, 2026

SECURE: Stable Early Collision Understanding via Robust Embeddings in Autonomous Driving

Authors: Wenjing Wang, Wenxuan Wang, Songning Lai

Organizations: Xiamen University Malaysia · Xinjiang University · HKUST(GZ)

Abstract

While deep learning has significantly advanced accident anticipation, the robustness of these safety-critical systems against real-world perturbations remains a major challenge. We reveal that state-of-the-art models like CRASH, despite their high performance, exhibit significant instability in predictions and latent representations when faced with minor input perturbations, posing serious reliability risks. To address this, we introduce SECURE - Stable Early Collision Understanding Robust Embeddings, a framework that formally defines and enforces model robustness. SECURE is founded on four key attributes: consistency and stability in both prediction space and latent feature space. We propose a principled training methodology that fine-tunes a baseline model using a multi-objective loss, which minimizes divergence from a reference model and penalizes sensitivity to adversarial perturbations. Experiments on DAD and CCD datasets demonstrate that our approach not only significantly enhances robustness against various perturbations but also improves performance on clean data, achieving new state-of-the-art results.

Figures & tables

Figure 1: The CRASH model exhibits significant output differences between clean and perturbed inputs.

Model	DAD [ 3 ]		CCD [ 1 ]
	AP(%) $\uparrow$	mTTA(s) $\uparrow$	AP(%) $\uparrow$	mTTA(s) $\uparrow$
DSA [ 3 ]	48.1	1.34	99.6	4.53
L-RAI [ 25 ]	51.4	3.01	98.9	3.32
AdaLEA [ 18 ]	62.3	2.94	99.2	3.45
DSTA [ 9 ]	59.2	2.60	98.8	4.53
UString [ 1 ]	58.0	2.85	99.5	4.52

Table 1: Comparison of model performance on the complete datasets. Bold and underlined values represent the best and second-best performance in each category.

Model / Dataset	Clean		IP (0.1)		IP (0.2)		LP (0.1)		LP (0.2)
Model / Dataset	AP(%) $\uparrow$	mTTA(s) $\uparrow$	AP(%) $\uparrow$	mTTA(s) $\uparrow$	AP(%) $\uparrow$	mTTA(s) $\uparrow$	AP(%) $\uparrow$	mTTA(s) $\uparrow$	AP(%) $\uparrow$	mTTA(s) $\uparrow$
CRASH (DAD)	67.21	3.020	65.83	3.053	63.37	2.964	64.41	2.984	44.73	3.119
SECURE (DAD)	68.67	3.127	68.42	3.120	68.32	3.062	68.19	3.083	57.61	3.207
CRASH (CCD)	99.53	4.554	99.29	4.540	99.17	3.965	98.88	3.074	69.26	3.413
SECURE (CCD)	99.73	4.583	99.67	4.483	99.48	4.481	99.29	3.191	81.45	3.430

Table 2: Comparison of CRASH and SECURE on DAD and CCD datasets under different perturbations. The number in the parentheses is the standard deviation of Gaussian noise.

$\mathcal{L}_{\text{Task}}$	$\mathcal{L}_{C_{ps}}$	$\mathcal{L}_{S_{pd}}$	$\mathcal{L}_{C_{lm}}$	$\mathcal{L}_{S_{ld}}$	AP(%) $\uparrow$	mTTA(s) $\uparrow$
$\checkmark$	$\times$	$\times$	$\times$	$\times$	63.37	2.964
$\checkmark$	$\checkmark$	$\times$	$\times$	$\times$	63.47	3.036
$\checkmark$	$\checkmark$	$\checkmark$	$\times$	$\times$	64.80	3.015
$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\times$	65.29	3.001
$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	68.32	3.062

Table 3: Ablation Results of SECURE on the DAD Dataset under IP (0.2)

Traffic Accident Anticipation

Traffic accident anticipation predicts imminent collisions or rising crash risk early enough to support driver warnings and proactive road-safety decisions.

See all 22 papers

Task-Aligned Stability Analysis of Vision-Language Models for Autonomous Driving Hazard Detection
Jun 10, 2026Everett RichardsVLM Evaluation VLM Robustness
Department of Computer Science, San Diego State University.
TrajRS: Towards Certified Robustness in Pedestrian Trajectory Prediction
Jun 27, 2026Liang Zhang, Gaojie Jin, Yao Shi +4Randomized Smoothing Pedestrian Trajectory Prediction
Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, China · University of Chinese Academy of Sciences, China · Hangzhou Institute for Advanced Study, UCAS, China +1
Evaluating randomized smoothing as a defense against adversarial attacks in trajectory prediction
Mar 11, 2026Julian F. Schumann, Eduardo Figueiredo, Frederik Baymler Mathiesen +3Probabilistic Trajectory Prediction Adversarial Robustness
Department of Cognitive Robotics, TU Delft, Delft, Netherlands · Delft Center for Systems and Control, TU Delft, Delft, Netherlands

Jun 10, 2026J/K move · Enter open · S save

Task-Aligned Stability Analysis of Vision-Language Models for Autonomous Driving Hazard Detection

Everett Richards

Department of Computer Science, San Diego State University.

VLM Evaluation VLM Robustness VLMs for Autonomous Driving Autonomous Driving Safety Evaluation Autonomous Driving Perception

Open paper AskPDF arXiv

Vision-language models (VLMs) are increasingly used for scene understanding in autonomous driving, but robustness analysis often relies on task-agnostic embedding stability alone. We study whether corruption-induced embedding drift predicts changes in a task-aligned hazard score derived from CLIP image-text similarities. Using controlled corruptions on BDD100K road scenes, we compare embedding drift against margin drift, defined as the change in hazard score under perturbation. The relationship is highly corruption-dependent: some families exhibit strong coupling between representation drift and decision drift, while others induce hazardous decision instability despite relatively modest embedding change. Furthermore, corruption families differ in failure direction: most suppress hazard detections via false negatives, while occlusion instead triggers false alarms, suggesting that benchmark design should account for asymmetric failure modes, not just overall instability rates. These results suggest that robustness benchmarks should include task-aligned stability measures in addition to embedding-level perturbation statistics.

Abstract

Traffic Accident Anticipation

Explore similar work

Task-Aligned Stability Analysis of Vision-Language Models for Autonomous Driving Hazard Detection

TrajRS: Towards Certified Robustness in Pedestrian Trajectory Prediction

Evaluating randomized smoothing as a defense against adversarial attacks in trajectory prediction