Scholarlink

This paper investigates the potential of PaliGemma, a recently released open-source Visual-Language Model (VLM), for detecting phone-based driver distraction. We evaluate PaliGemma’s performance on the State Farm Distracted Driver Detection (SFDDD) dataset. The proposed method involves three steps: (1) generating a textual description of the input image using PaliGemma, (2) create prompts related to phone distraction based on the text description (e.g., “Is the driver distracted by talking or using the phone?”), and (3) identifying the driver’s phone use status (talking or texting) based on PaliGemma’s response. We compare the accuracy of three PaliGemma models for classifying four driver actions: texting/talking on the phone with left/right hand. All models achieved high accuracy (>95%) for most categories, with PaliGemma-3b-pt-896 achieving impressive 100% accuracy. Our findings suggest that PaliGemma VLMs offer promising potential for detecting phone distraction while driving.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detection of Phone Distraction While Driving Using Open Visual-Language Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Florin Girbacia,
  • Gheorghe Daniel Voinea,
  • Manuela Daniela Danu,
  • Ioana-Diana Buzdugan,
  • Mihai Duguleana

摘要

This paper investigates the potential of PaliGemma, a recently released open-source Visual-Language Model (VLM), for detecting phone-based driver distraction. We evaluate PaliGemma’s performance on the State Farm Distracted Driver Detection (SFDDD) dataset. The proposed method involves three steps: (1) generating a textual description of the input image using PaliGemma, (2) create prompts related to phone distraction based on the text description (e.g., “Is the driver distracted by talking or using the phone?”), and (3) identifying the driver’s phone use status (talking or texting) based on PaliGemma’s response. We compare the accuracy of three PaliGemma models for classifying four driver actions: texting/talking on the phone with left/right hand. All models achieved high accuracy (>95%) for most categories, with PaliGemma-3b-pt-896 achieving impressive 100% accuracy. Our findings suggest that PaliGemma VLMs offer promising potential for detecting phone distraction while driving.