AI-Enabled Lightweight Video Stream Processing for Electric Shock Prevention Monitoring in Smart Grid Edge Scenarios

INTRODUCTION: Real-time early warning of electric shock risks is urgently needed in edge scenarios of smart grids. Existing visual methods are still limited by scene adaptability, end-to-end lightweight design, and insufficient field verification. Objective: The objective of the study is to propose a lightweight video stream processing algorithm for power operation sites, balancing accuracy, latency, and deployment efficiency on edge devices. METHODS: A constrained multi-objective model prioritizing safety risks is constructed, integrating risk-weighted ROI extraction, adaptive inter-frame filtering, lightweight image enhancement, ROI-aware YOLOv8n, pruning, and pipelined inference. RESULTS: Data analysis shows that the model has 2.86M parameters, 3.12 GFLOPs of computation, a mAP@0.5 of 93.27%, and a Jetson Nano speed of 31.4 fps. In substations, the average mAP@0.5 is 92.14%, the early warning latency is 126 ms, and the false negative rate in rain, fog, low light, and occlusion scenarios is approximately 1%. CONCLUSION: This method reduces redundant computation while maintaining high accuracy and low latency, providing an edge-intelligent solution for electric shock protection in smart grids.