Artificial intelligence–assisted occupational safety modeling

In this thesis, a proactive approach to accident prevention in the field of Occupational Health and Safety (OHS) is proposed through an artificial intelligence (AI)-based modeling framework. Traditional OHS methods are typically reactive, with interventions occurring after accidents have taken place. The model presented here aims to enable real-time detection of potential hazards and unsafe conditions in workplaces and to prevent possible accidents in advance. The primary objective of the study is to develop a computer vision system that automatically monitors the use of personal protective equipment (PPE). To this end, an object detection model was built using the "You Only Look Once (YOLOv5)" algorithm. For model training, a dataset of 2,512 images containing the classes "person," "helmet," "vest," and "glove" was assembled. The images were split into training (70%), validation (20%), and test (10%) sets and subjected to an annotation process. This process constitutes the foundational step that enables the system to learn from visual inputs and perform accurate classifications. Training was conducted for 100 epochs, targeting the reduction of loss functions and the improvement of mean Average Precision (mAP) metrics. The results indicate that the model achieved an mAP@0.5 of 86.6%. For the helmet (97.8%) and vest (96.5%) classes, both precision and recall exceeded 97%, demonstrating strong detection performance for these categories. However, glove detection (56.5%) was comparatively weaker, likely due to the small size of the objects, their diverse types, and frequent occlusions. These findings clearly reveal the model's practical limitations and the areas requiring improvement. The developed system can be integrated into existing security camera infrastructures to detect PPE violations in real time. Diversifying the dataset—by incorporating different lighting conditions and additional PPE types—may further enhance accuracy. Moreover, the system's ethical dimensions (e.g., safeguarding employee privacy and addressing potential legal liabilities stemming from false detections) should be examined in detail. Future field tests and sector-specific adaptations will represent important steps toward achieving the "zero-accident" goal.