Video Data-Driven Approaches for Monitoring Hazard Sources in Emergency Response Scenarios at Coal and Renewable Energy Facilities
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The rapid expansion of coal and renewable energy enterprises has introduced complex, heterogeneous visual environments-ranging from low-light coal mines to high-glare photovoltaic stations-which pose significant challenges for safety monitoring. Conventional surveillance strategies often fail to meet intrinsic safety requirements due to poor environmental adaptability and the “long-tail” scarcity of accident samples. To address these limitations, this study proposes a Full-Transformer MultiModal Risk Perception Framework. Unlike loose assemblies of independent models, we construct a unified collaborative system that synergizes static detection, dynamic behavior recognition, and anomaly identification. For static perception, we integrate adaptive illumination calibration and an Asymptotic Feature Pyramid Network (AFPN) into RT-DETR to robustly handle occlusions and small targets specific to industrial scenes. For dynamic behavior, a Semantic-Aware Masked Spatiotemporal Modeling strategy is employed to enhance violation recognition efficiency. Furthermore, we design a Multi-Scale Local Manifold Estimation model based on FastFlow to detect unknown anomalies without over-generalization. A key conceptual innovation, the Bidirectional Cross-Attention Fusion (Bi-MCAF) module, semantically aligns these heterogeneous features to output interpretable risk levels and causality. Validated on our “Energy MultiScenario Safety Dataset (EMS-Dataset),” the method achieves an mAP@ 0.5 of 86.4% (static), Top-1 Accuracy of 93.1% (dynamic), and AUROC of 97.3% (anomaly). These results significantly surpass YOLOv5 and CNN-LSTM baselines, demonstrating the framework's potential for reliable hazard monitoring in complex energy facilities.
Fields of Study & Subjects
Advanced Neural Network Applications
Video Surveillance and Tracking Methods
Human Pose and Action Recognition
Publication details
- Journal / source
- International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
- Pages
- 722–725
- Publication date
- 16 January 2026
- Language
- EN