Scholarlink

<p>Real-time object detection is essential for autonomous driving, as it enables rapid decision-making in dynamically changing environments. Among recent advancements, <i>fully convolutional one-stage with a single head (FCOSH)</i> object detector has emerged as a simplified <i>single-head</i> version of <i>FCOS</i>. Despite its computational efficiency, <i>FCOSH</i> falls short of real-time deployment standards. A widely used strategy to improve inference speed is input image downsizing. While effective in reducing computational load, it causes <i>FCOSH</i> to suffer from spatial information loss, resulting in degraded detection accuracy. To bridge this gap, we propose confining the use of high-resolution feature maps to the final prediction stage. This avoids the overhead of processing high resolution feature maps throughout the entire network while restoring spatial precision only where it is most impactful. At the core of this approach, we propose the <i>spatially-focused upsampling block (SFUB)</i>, a novel lightweight module that reconstructs high-resolution feature maps from coarse representations immediately prior to the prediction layers. The resulting architecture, <i>RT-FCOSH</i>, is a real-time variant of <i>FCOSH</i> designed for autonomous driving. Experiments on BDD100K and TJU-DHD datasets using inputs downscaled to 60% and 40% show that RT-FCOSH achieves 26.2% AP at 45 FPS on BDD100K (+2.6% over FCOSH) and 49.0% AP at 47 FPS on TJU-DHD (+3.7%), confirming its effectiveness under low-resolution constraints</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

RT-FCOSH: bridging accuracy and efficiency in low-resolution object detection for autonomous driving

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Saad Mboutayeb,
  • Aicha Majda,
  • Khalid Zenkouar

摘要

Real-time object detection is essential for autonomous driving, as it enables rapid decision-making in dynamically changing environments. Among recent advancements, fully convolutional one-stage with a single head (FCOSH) object detector has emerged as a simplified single-head version of FCOS. Despite its computational efficiency, FCOSH falls short of real-time deployment standards. A widely used strategy to improve inference speed is input image downsizing. While effective in reducing computational load, it causes FCOSH to suffer from spatial information loss, resulting in degraded detection accuracy. To bridge this gap, we propose confining the use of high-resolution feature maps to the final prediction stage. This avoids the overhead of processing high resolution feature maps throughout the entire network while restoring spatial precision only where it is most impactful. At the core of this approach, we propose the spatially-focused upsampling block (SFUB), a novel lightweight module that reconstructs high-resolution feature maps from coarse representations immediately prior to the prediction layers. The resulting architecture, RT-FCOSH, is a real-time variant of FCOSH designed for autonomous driving. Experiments on BDD100K and TJU-DHD datasets using inputs downscaled to 60% and 40% show that RT-FCOSH achieves 26.2% AP at 45 FPS on BDD100K (+2.6% over FCOSH) and 49.0% AP at 47 FPS on TJU-DHD (+3.7%), confirming its effectiveness under low-resolution constraints