BD-HazardVLM: Probing Vision-Language Models for Latent Defensive-Driving Hazards in Bangladesh Road Scenes

Vision-language models are increasingly used for driving-scene understanding, yet their defensive reasoning under weakly structured Global South traffic remains unclear.

视觉语言模型正越来越多地被用于驾驶场景理解，然而它们在结构松散的全球南方交通环境下的防御性推理能力尚不明确。

We introduce BD-HazardVLM, a diagnostic benchmark of 500 Bangladesh road-scene images, comprising 375 from RSUD20K and 125 from TFP-BD, annotated for hazard presence, primary hazard category, risk timing, severity, defensive actions, visible evidence, and hallucination- and Western-road-bias traps.

我们引入了 BD-HazardVLM，这是一个包含 500 张孟加拉国道路场景图像的诊断基准，其中 375 张来自 RSUD20K，125 张来自 TFP-BD。该基准标注了危险存在情况、主要危险类别、风险时机、严重程度、防御性动作、可见证据，以及幻觉和西方道路偏见陷阱。

We evaluate Qwen2.5-VL-7B, InternVL3-8B, and MiniCPM-V-4.6 using a shared zero-shot task definition and structured output schema. Binary hazard-presence accuracy ranges from 88.0 percent to 89.2 percent, essentially matching the 89.0 percent majority-class baseline.

我们使用统一的零样本任务定义和结构化输出模式，对 Qwen2.5-VL-7B、InternVL3-8B 和 MiniCPM-V-4.6 进行了评估。二元危险存在检测的准确率在 88.0% 到 89.2% 之间，基本与 89.0% 的多数类基准持平。

Fine-grained reasoning is substantially weaker: category accuracy remains below 40 percent, risk-timing accuracy ranges from 1.6 percent to 12.4 percent, and exact defensive-action matching remains below 2 percent.

细粒度推理能力则明显较弱：类别准确率保持在 40% 以下，风险时机准确率在 1.6% 到 12.4% 之间，而精确的防御性动作匹配率则低于 2%。

Although InternVL3-8B achieves the highest severity accuracy at 61.2 percent, it has zero recall on the 58 high-severity cases. Qwen exhibits severe category mode collapse, while all models show weak latent-hazard anticipation and near-zero self-reported trap recognition.

尽管 InternVL3-8B 在严重程度判断上达到了 61.2% 的最高准确率，但它对 58 个高严重性案例的召回率为零。Qwen 表现出严重的类别模式崩溃，而所有模型在潜在危险预判方面均表现较弱，且对陷阱的自我识别能力几乎为零。

For the evaluated models, these findings show that coarse hazard recognition does not imply reliable fine-grained defensive-driving reasoning and motivate geographically diverse safety evaluation.

对于所评估的模型而言，这些发现表明粗略的危险识别并不意味着可靠的细粒度防御性驾驶推理，这也促使我们需要进行地理多样化的安全性评估。