Scholarlink

<p>Detecting faint, camouflaged anthropogenic traces (such as footprints, disturbed soil) in complex outdoor environments is a critical yet challenging task for vision-based robotic systems in search and rescue missions. While deep learning has advanced camouflaged object detection (COD), state-of-the-art methods rely heavily on large-scale pixel-level annotations and offline training, rendering them brittle in real-world field operations where targets are non-continuous, labels are absent, and image quality degrades due to motion blur and uneven lighting. To address these engineering challenges, this paper proposes a self-supervised machine vision framework that enables real-time online adaptation on robotic platforms. The approach integrates a two-stage learning strategy: it first acquires general COD knowledge from public datasets via supervised pre-training, then seamlessly switches to an online self-supervised fine-tuning phase during deployment. This phase is guided by two core innovations designed for robust vision in the field: a texture difference loss based on Gray-Level Co-occurrence Matrix (GLCM) contrast to capture subtle target-background visual discrepancies, and a position stability loss to ensure temporal consistency of predictions across video frames, together forming a self-supervision signal that requires no manual annotations. Experiments demonstrate that our method effectively segments camouflaged traces in low-quality, unlabeled outdoor images captured by a mobile robot, outperforming existing COD models in adaptation speed and segmentation accuracy on a challenging custom rescue dataset. Notably, it maintains a low false-alarm rate (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(0.88\%\)</EquationSource> </InlineEquation>) in negative scenes and runs in real time on the robot’s onboard hardware, confirming its practical suitability for deployed vision systems. This work advances machine vision for field robotics by enabling online learning under degraded imaging conditions with minimal human intervention.</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

Real-time detection of outdoor non-obvious anthropogenic trace via texture contrast learning

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Shuaiqi Wang,
  • Jiajie Sha,
  • Wei You,
  • Yuanxiang Wang,
  • Qirong Tang

摘要

Detecting faint, camouflaged anthropogenic traces (such as footprints, disturbed soil) in complex outdoor environments is a critical yet challenging task for vision-based robotic systems in search and rescue missions. While deep learning has advanced camouflaged object detection (COD), state-of-the-art methods rely heavily on large-scale pixel-level annotations and offline training, rendering them brittle in real-world field operations where targets are non-continuous, labels are absent, and image quality degrades due to motion blur and uneven lighting. To address these engineering challenges, this paper proposes a self-supervised machine vision framework that enables real-time online adaptation on robotic platforms. The approach integrates a two-stage learning strategy: it first acquires general COD knowledge from public datasets via supervised pre-training, then seamlessly switches to an online self-supervised fine-tuning phase during deployment. This phase is guided by two core innovations designed for robust vision in the field: a texture difference loss based on Gray-Level Co-occurrence Matrix (GLCM) contrast to capture subtle target-background visual discrepancies, and a position stability loss to ensure temporal consistency of predictions across video frames, together forming a self-supervision signal that requires no manual annotations. Experiments demonstrate that our method effectively segments camouflaged traces in low-quality, unlabeled outdoor images captured by a mobile robot, outperforming existing COD models in adaptation speed and segmentation accuracy on a challenging custom rescue dataset. Notably, it maintains a low false-alarm rate ( \(0.88\%\) ) in negative scenes and runs in real time on the robot’s onboard hardware, confirming its practical suitability for deployed vision systems. This work advances machine vision for field robotics by enabling online learning under degraded imaging conditions with minimal human intervention.