Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluation of Safety Cognition Capability in Vision-Language Models for Autonomous Driving

Enming Zhang, Peizhe Gong, Xingyuan Dai, Min Huang, Yisheng Lv, Qinghai Miao

TL;DR

This work tackles the safety-cognition alignment problem for vision-language models in autonomous driving by introducing SCD-Bench, a four-dimension safety evaluation framework, and ADA for scalable annotation. It further provides SCD-Training, a large-scale dataset to train safety-aware models, and an automated GPT-4o-based evaluation pipeline achieving high agreement with human judgments. Empirical results show existing open models underperform on safety cognition (SR < 60%), while SCD-Model-1B trained with SCD-Training attains around 90% safety on SCD-Bench, and general-task improvements accompany domain-specific gains. The framework offers a practical path toward robust, safety-aligned VLMs for interactive autonomous-driving scenarios with strong cross-domain benefits.

Abstract

Ensuring the safety of vision-language models (VLMs) in autonomous driving systems is of paramount importance, yet existing research has largely focused on conventional benchmarks rather than safety-critical evaluation. In this work, we present SCD-Bench (Safety Cognition Driving Benchmark) a novel framework specifically designed to assess the safety cognition capabilities of VLMs within interactive driving scenarios. To address the scalability challenge of data annotation, we introduce ADA (Autonomous Driving Annotation), a semi-automated labeling system, further refined through expert review by professionals with domain-specific knowledge in autonomous driving. To facilitate scalable and consistent evaluation, we also propose an automated assessment pipeline leveraging large language models, which demonstrates over 98% agreement with human expert judgments. In addressing the broader challenge of aligning VLMs with safety cognition in driving environments, we construct SCD-Training, the first large-scale dataset tailored for this task, comprising 324.35K high-quality samples. Through extensive experiments, we show that models trained on SCD-Training exhibit marked improvements not only on SCD-Bench, but also on general and domain-specific benchmarks, offering a new perspective on enhancing safety-aware interactions in vision-language systems for autonomous driving.

Evaluation of Safety Cognition Capability in Vision-Language Models for Autonomous Driving

TL;DR

This work tackles the safety-cognition alignment problem for vision-language models in autonomous driving by introducing SCD-Bench, a four-dimension safety evaluation framework, and ADA for scalable annotation. It further provides SCD-Training, a large-scale dataset to train safety-aware models, and an automated GPT-4o-based evaluation pipeline achieving high agreement with human judgments. Empirical results show existing open models underperform on safety cognition (SR < 60%), while SCD-Model-1B trained with SCD-Training attains around 90% safety on SCD-Bench, and general-task improvements accompany domain-specific gains. The framework offers a practical path toward robust, safety-aligned VLMs for interactive autonomous-driving scenarios with strong cross-domain benefits.

Abstract

Ensuring the safety of vision-language models (VLMs) in autonomous driving systems is of paramount importance, yet existing research has largely focused on conventional benchmarks rather than safety-critical evaluation. In this work, we present SCD-Bench (Safety Cognition Driving Benchmark) a novel framework specifically designed to assess the safety cognition capabilities of VLMs within interactive driving scenarios. To address the scalability challenge of data annotation, we introduce ADA (Autonomous Driving Annotation), a semi-automated labeling system, further refined through expert review by professionals with domain-specific knowledge in autonomous driving. To facilitate scalable and consistent evaluation, we also propose an automated assessment pipeline leveraging large language models, which demonstrates over 98% agreement with human expert judgments. In addressing the broader challenge of aligning VLMs with safety cognition in driving environments, we construct SCD-Training, the first large-scale dataset tailored for this task, comprising 324.35K high-quality samples. Through extensive experiments, we show that models trained on SCD-Training exhibit marked improvements not only on SCD-Bench, but also on general and domain-specific benchmarks, offering a new perspective on enhancing safety-aware interactions in vision-language systems for autonomous driving.

Paper Structure

This paper contains 32 sections, 3 equations, 8 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Vision-Language Models
  4. Benchmarking VLMs in Autonomous Driving
  5. Safety-Cognitive Driving Benchmark
  6. SCD-Bench Tasks
  7. Command Misunderstanding
  8. Malicious Decision
  9. Perception Induction
  10. Ethical Dilemmas
  11. Data Construction
  12. Data Collection
  13. Annotation
  14. Evaluation Method
  15. Criterion
  16. ...and 17 more sections

Figures (8)

  • Figure 1: The SCD-Benchmark Overview. Figure (a) presents the four types of tasks in the SCD-Benchmark. Figure (b) shows the performance of closed-source and open-source models on the SCD-Benchmark.
  • Figure 2: Categories with fine-grained dimensions and their corresponding examples of SCD-Bench.
  • Figure 3: The overall framework of the Autonomous Driving Image-Text Annotation System.
  • Figure 4: The Distribution of Image Sources in SCD-Bench.
  • Figure 5: The 2D Bounding Box Distribution of SCD-Bench.
  • ...and 3 more figures