Scholarlink

<p>Remote monitoring of fully mechanized longwall faces requires operators to recognize key targets, particularly the shearer and personnel, and to interpret their positions in relation to the longwall-face structure. Split-screen displays force operators to infer cross-view relationships mentally, increasing cognitive load and the risk of delayed or incorrect safety decisions. Direct video stitching offers a unified display but is unreliable in longwall scenes because moving foregrounds, dust, vibration, low visibility, and large inter-view perspective differences cause registration drift and ghosting. This paper proposes a background-foreground decoupled visualization framework under a unified spatial reference. Its main novelty lies in separating reference-plane construction from dynamic-foreground rendering: stable background structures are used to establish the spatial coordinate system, whereas key targets are segmented, anchored, and reprojected as independent foreground overlays. A single-view selection strategy is then used in overlap regions to suppress duplicate target imagery. Experiments using multi-camera video from the 22207 longwall face of Yaping Coal Mine produced an average reprojection RMSE of 2.79 pixels, an end-to-end latency of 128 ms at 27.6 fps, an instance-segmentation mAP@0.5 of 0.941 for shearer and personnel targets, and a reduction in ghosting rate from 18.6% to 1.9% compared with SIFT-based stitching. Although based on a single longwall-face configuration, the field results demonstrate feasibility in the tested deployment setting. Simulated visual-degradation tests further indicate stable behavior under moderate visibility loss, while performance under peak cutting-dust conditions requires further underground validation. Overall, the framework provides a spatially consistent, target-centered overview to support safer remote monitoring and more timely operational decisions in dynamic longwall environments.</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

Key-target visualization in longwall faces via background–foreground decoupling under a unified spatial reference

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月2日

查看全文
  • Zhenping Li,
  • Sanzhen Wang,
  • Qing Liu,
  • Dongqiang Wang,
  • Chunlei Yin,
  • Xinglong Huang,
  • Fengtao Chen,
  • Jieliang Yang

摘要

Remote monitoring of fully mechanized longwall faces requires operators to recognize key targets, particularly the shearer and personnel, and to interpret their positions in relation to the longwall-face structure. Split-screen displays force operators to infer cross-view relationships mentally, increasing cognitive load and the risk of delayed or incorrect safety decisions. Direct video stitching offers a unified display but is unreliable in longwall scenes because moving foregrounds, dust, vibration, low visibility, and large inter-view perspective differences cause registration drift and ghosting. This paper proposes a background-foreground decoupled visualization framework under a unified spatial reference. Its main novelty lies in separating reference-plane construction from dynamic-foreground rendering: stable background structures are used to establish the spatial coordinate system, whereas key targets are segmented, anchored, and reprojected as independent foreground overlays. A single-view selection strategy is then used in overlap regions to suppress duplicate target imagery. Experiments using multi-camera video from the 22207 longwall face of Yaping Coal Mine produced an average reprojection RMSE of 2.79 pixels, an end-to-end latency of 128 ms at 27.6 fps, an instance-segmentation mAP@0.5 of 0.941 for shearer and personnel targets, and a reduction in ghosting rate from 18.6% to 1.9% compared with SIFT-based stitching. Although based on a single longwall-face configuration, the field results demonstrate feasibility in the tested deployment setting. Simulated visual-degradation tests further indicate stable behavior under moderate visibility loss, while performance under peak cutting-dust conditions requires further underground validation. Overall, the framework provides a spatially consistent, target-centered overview to support safer remote monitoring and more timely operational decisions in dynamic longwall environments.