Construction sites consistently rank among the most hazardous occupational environments worldwide, with head injuries from falling or flying objects identified as a primary contributor to construction fatalities in every major market. Despite the universal regulatory mandate for safety helmet usage, non-compliance remains pervasive owing to the practical impossibility of maintaining continuous manual supervision across large, complex sites. Traditional automated monitoring approaches based on conventional computer vision techniques have demonstrated insufficient accuracy for reliable deployment in the visually complex conditions typical of active construction environments, while commercial AIbased platforms impose subscription costs prohibitive to small and medium contractors.This paper presents SafeHelmet Vision AI, a deep learning-based industrial safety monitoring system designed to automate helmet compliance detection at construction sites and related industrial workplaces. The proposed system employs a YOLOv8n (nano) object detection model trained via transfer learning from COCO-pretrained weights on a domain-specific dataset of 4,200 annotated construction site images encompassing 9,800 helmet and 8,200 worker bounding box instances. Training was conducted for 100 epochs with AdamW optimisation (lr = 0.001), comprehensive data augmentation including mosaic, HSV perturbation, random flip, rotation (10), and scale variation (50%), yielding a validation mean Average Precision at IoU = 0.50 (mAP50) of 97.8%, a precision of 95.2%, and a recall of 94.9%.The trained model is integrated into a Streamlit web application that accepts uploaded construction site images in JPG, PNG, or BMP formats and returns annotated detection results with bounding boxes, class labels, and confidence scores within a mean inference latency of 165 milliseconds on standard CPU hardware. An automated safety compliance assessment engine evaluates helmet-toperson count ratios and generates colour-coded violation alerts. Comparative evaluation demonstrates a 27.2 percentage-point precision advantage and a 33.9 percentage-point recall advantage over a traditional Haar cascade baseline. The complete system requires no client-side installation and is deployable to Streamlit Cloud from a GitHub repository with a single configuration step, making enterprise-grade safety monitoring accessible to safety personnel without specialised technical training..
Vani et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: