Research on large-scale visual models for power systems in scenarios requiring detailed monitoring for safe operations

Addressing the issues of low monitoring accuracy and insufficient robustness in safety operation monitoring within power utility work scenarios, this paper proposes a multi-view safety helmet detection method based on hierarchical heterogeneous graph neural networks (HGNN). This approach integrates video foreground segmentation with multi-level feature learning to achieve refined recognition of workers’ protective behaviors in a complex environments. First, a Gaussian Mixture Model (GMM) performs background statistical modeling and dynamic updates on video pixels, enabling adaptive separation of background and foreground to extract effective work areas. Subsequently, a sequential encoder and hierarchical attention mechanism embed multi-view images as graph nodes, constructing a hierarchical graph structure with cross-view and global connections to capture spatial relationships and contextual dependencies. The model achieves cross-view feature alignment through information aggregation and self-attention, significantly enhancing robustness against lighting variations, occlusions, and scale differences. Experiments on the Datafountain hardhat dataset demonstrate a 6% improvement in mean average precision (mAP) over baseline models like DenseNet, R-CNN, and Fast R-CNN, with an IoU detection accuracy of 96.41%. Variance and entropy analysis further validate the model's stability in distinguishing multi-view features. Class activation map (CAM) visualizations demonstrate the model's ability to accurately focus on helmet regions, exhibiting strong interpretability.