Research on Substation Personnel Fall Detection Method Based on Vision-Language Models

In power substation environments, ensuring personnel safety is critical to maintaining stable and reliable power operations. Detecting falls—a key safety concern—involves overcoming challenges such as low lighting, occlusion by equipment, and visually cluttered backgrounds. Conventional vision-only methods often fail to provide robust performance under these real-world conditions. In this work, we propose a novel fall detection approach leveraging large-scale pre-trained Vision-Language Models (VLMs). By integrating textual semantics describing "fall" incidents with visual features from substation surveillance footage and performing incremental fine-tuning on domain-specific annotated datasets, our method achieves high-precision detection under challenging scenarios. Experimental results demonstrate that the proposed approach surpasses traditional CNN-based methods in terms of accuracy, robustness, and generalization. The findings offer a promising solution for intelligent safety monitoring and early warning systems in the power industry, reducing risks associated with personnel injuries and operational disruptions.