Abstract: To address the problems of single-frame false detections, occlusion-induced missed detections, and delayed warning responses in safety helmet wearing detection at power cable construction sites, this study investigates an early-warning approach for construction personnel based on an improved YOLOv8n model. First, the input video stream is subjected to dynamic cropping of the power cable construction area and illumination-adaptive enhancement, thereby eliminating irrelevant background interference and improving image quality. Subsequently, an improved YOLOv8n model is employed to perform real-time detection and recognition of workers’ head regions and helmet-wearing status within the cropped area, distinguishing among three conditions, namely, helmet worn, helmet not worn, and head occluded. High-confidence detection results are then output as the basis for early-warning decisions. Furthermore, a temporal voting early-warning module based on a sliding queue is designed to accumulate confidence scores across consecutive frames and conduct dynamic-threshold decision-making, thereby triggering graded audible and visual alarms, recording violation images, and pushing alarm information to the safety supervision platform of the power construction site. Comparative experiments demonstrate that the proposed method outperforms conventional approaches in detection accuracy, robustness, and warning timeliness, and can effectively satisfy the engineering application requirements of intelligent supervision of safety helmet use in power cable construction.