Scholarlink

<p>Ensuring miner safety in underground environments demands vision systems that combine high accuracy with real-time performance under resource constraints. While existing video-based monitoring solutions suffer from limitations where CNNs struggle with long-range modeling and Transformers face high computational complexity, we propose MinerUBR, a compact skeleton sensing framework that converts raw video streams into compact 3D heatmap volumes via pose estimation, replacing traditional graph sequence inputs. By leveraging a novel bidirectional spatiotemporal compression module based on video state space modeling, enhanced via bidirectional scanning, our method achieves efficient joint modeling of skeleton sequences with only 1.9M parameters while retaining spatiotemporal information through spatial and temporal embeddings. Validated on both public datasets including NTURGB+D and our self-constructed dataset featuring 20 unsafe miner behaviors collected using Kinect V2 sensors in simulated mine environments, MinerUBR-B achieves 91.2% accuracy on NTU60 and 92.8% accuracy on the mining dataset with only 1.9M parameters. The system operates at 138 FPS on CPU platforms, demonstrating 10.6 times and 3.8 times faster inference speed compared to PoseC3D and ST-GCN, respectively. This work provides practical insights for implementing lightweight behavior recognition in resource-constrained mining environments, balancing accuracy, and efficiency for industrial deployment.</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 recognition of miner unsafe behaviors via skeleton-based spatiotemporal modeling with Mamba

  • 期刊论文

  • 发表日期：

    2025年9月17日

查看全文
  • Biao Li,
  • Shoufeng Tang,
  • Wenyi Li

摘要

Ensuring miner safety in underground environments demands vision systems that combine high accuracy with real-time performance under resource constraints. While existing video-based monitoring solutions suffer from limitations where CNNs struggle with long-range modeling and Transformers face high computational complexity, we propose MinerUBR, a compact skeleton sensing framework that converts raw video streams into compact 3D heatmap volumes via pose estimation, replacing traditional graph sequence inputs. By leveraging a novel bidirectional spatiotemporal compression module based on video state space modeling, enhanced via bidirectional scanning, our method achieves efficient joint modeling of skeleton sequences with only 1.9M parameters while retaining spatiotemporal information through spatial and temporal embeddings. Validated on both public datasets including NTURGB+D and our self-constructed dataset featuring 20 unsafe miner behaviors collected using Kinect V2 sensors in simulated mine environments, MinerUBR-B achieves 91.2% accuracy on NTU60 and 92.8% accuracy on the mining dataset with only 1.9M parameters. The system operates at 138 FPS on CPU platforms, demonstrating 10.6 times and 3.8 times faster inference speed compared to PoseC3D and ST-GCN, respectively. This work provides practical insights for implementing lightweight behavior recognition in resource-constrained mining environments, balancing accuracy, and efficiency for industrial deployment.