Scholarlink

<p>Vehicle detection under hostile driving conditions remains a significant challenge, particularly in adverse scenarios such as low-light, snowy, rain, or night conditions. Therefore, this research proposes a novel detection framework that integrates a Disparity-guided Obstacle Attention Module (DOAM) into the backbone and a Multiscale and Efficient Channel Attention (MECA) module into the detection neck. The DOAM studies depth features to guide the network’s focus on obstacle regions only, while the MECA improves feature discrimination across various scales and channels. Comprehensive experiments conducted on the KITTI, CCD, and HCI datasets demonstrate that our method achieves better performance compared to state-of-the-art approaches, reaching an mAP@50–95 of 61.35% at night and 65.57% during the day, representing a 4–5% mAP improvement over the latest YOLOv12 baseline. Moreover, the proposed approach maintains a fast inference speed of 6.81 ms, highlighting its practicality for real-time deployment. These results underscore the effectiveness and efficiency of the proposed architecture, particularly in challenging real-world scenarios.</p>

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

DOAM-MECA: A Robust Framework for Vehicle Detection Under Adverse Driving Conditions

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Nhan Huu Tran,
  • Khoa Anh Dao,
  • Vinh Dinh Nguyen

摘要

Vehicle detection under hostile driving conditions remains a significant challenge, particularly in adverse scenarios such as low-light, snowy, rain, or night conditions. Therefore, this research proposes a novel detection framework that integrates a Disparity-guided Obstacle Attention Module (DOAM) into the backbone and a Multiscale and Efficient Channel Attention (MECA) module into the detection neck. The DOAM studies depth features to guide the network’s focus on obstacle regions only, while the MECA improves feature discrimination across various scales and channels. Comprehensive experiments conducted on the KITTI, CCD, and HCI datasets demonstrate that our method achieves better performance compared to state-of-the-art approaches, reaching an mAP@50–95 of 61.35% at night and 65.57% during the day, representing a 4–5% mAP improvement over the latest YOLOv12 baseline. Moreover, the proposed approach maintains a fast inference speed of 6.81 ms, highlighting its practicality for real-time deployment. These results underscore the effectiveness and efficiency of the proposed architecture, particularly in challenging real-world scenarios.