Computer vision-based recognition of spatial relevance between engineering vehicles and workers for monitoring struck-by accidents

Struck-by accidents remain a significant concern in construction safety, often resulting in severe injuries or fatalities. Recent advancements in computer vision technology present promising prospects for monitoring struck-by accidents at construction sites. However, practical deployment has been hindered by insufficient accuracy and the lack of methodologies for assessing collision risk levels after object detection. This study introduces a novel computer vision-based approach of spatial relevance for monitoring struck-by accidents at construction sites. The loss function was optimized by employing four intersection over union (IoU) variants to improve the accuracy. Following object detection, high-risk and dangerous spatial relevance were identified through virtual experiments. Field tests were conducted in different construction scenarios to validate the performance of the proposed method. The experiments results demonstrated a maximum improvement of 8.3% in mean average precision at 50% IoU (mAP50) under various environmental conditions. Field tests results achieved a high mAP50, 92.7%, and a low miss rate, 10.7%, thus enhancing struck-by hazards monitoring at construction sites.