Monitoring struck-by accidents in nighttime construction operations based on deep learning illumination enhancement and collision estimation

Purpose Struck-by accidents are the primary cause of fatalities in construction, accounting for over 75% of occupational deaths during nighttime operations. While computer vision techniques show promise for preventing such accidents, existing studies focus on daytime scenarios, neglecting critical nighttime challenges including severe illumination variations and limited visibility that degrade detection reliability. Design/methodology/approach To address these issues, this study presents a vision-based method for preventing struck-by accidents in nighttime construction operations. The proposed framework consists of four modules, including deep learning illumination enhancement, improved object detection, multi-object tracking and collision estimation. Findings Compared with the baseline YOLOv8 model without illumination enhancement and architectural improvement, experimental results indicated that the method achieved a 17.6%–28.1% improvement in detection accuracy under various lighting conditions. Field test results showed that the proposed method provided 2.08–3.51 seconds advance warnings with distance estimation errors ranging from 0.07–0.28 m under different lighting conditions. Originality/value The proposed method enables automated struck-by accidents prevention, significantly enhancing proactive safety management for 24/7 construction operations.