A Real-Time Safety Helmet Recognition Based on Yolov7

Despite the potential of safety helmets to help construction workers prevent accidents, they are often not worn correctly for several reasons. Nonetheless, the manual supervision of construction workers to guarantee the proper use of safety helmets is labor-intensive and inefficient. The evolution of target detection algorithms has progressed through the eras of manual feature generation and deep learning, with single-stage target detection algorithms outperforming others in terms of speed and performance. Despite advancements in safety helmet detection through deep learning algorithms, difficulties, including low accuracy and inadequate performance in challenging scenarios, remain. This project seeks to create a realtime safety helmet recognition system with YOLOv7, accompanied by Telegram notifications. The Hardhatwearing dataset from Harvard Dataverse, which consisted 7,035 images was used to train the YOLOv7 model under various conditions across 30 epochs. This study utilized three categories: 'helmet,' 'head,' and 'person.' Experimental results indicated that the YOLOv7 model in Experiment 6 achieved recall, [email protected], and [email protected] values of 0.95, 0.99, and 0.64, respectively. The model demonstrated improved accuracy in comparison to other YOLOv7 models.