Enhanced YOLOv5 algorithm for helmet wearing detection via combining bi-directional feature pyramid, attention mechanism and transfer learning

The complexity of infrastructure scenarios has led to a sustained increase in the global number of worksites related fatalities and injuries. Therefore, safety helmets play an essential role in protecting construction workers from accidents. It is essential to detect whether the helmets are correctly worn on the heads for smart construction site. However, due to the complex construction environments, it is challenging to precisely detect safety helmet wearing in real-time. This paper proposes an enhanced version of You Only Look Once version 5 (YOLOv5) to improve the detection accuracy, where bi-directional feature pyramid network (BiFPN), attention mechanism, and transfer learning are fully integrated. The BiFPN is taken to replace the original feature pyramid network (FPN) via adding additional cross layer edges with adaptive connecting weights. Attention mechanism is added after the end of backbone and neck network to let the network pay more attention on the interested region. Transfer learning is adopted for model training. The model is pre-trained by a head detection database and then fine-tuned by the helmet database. The proposed enhanced YOLOv5 is tested on a public GDU-HWD dataset, where both helmet and its color can be identified. This study achieves the accuracy at 93.3%, which is 4.8% higher than that of the original YOLOv5, but does not bring in much computing burden to the network. It is believed that the enhance version can also be successfully used in other similar detection tasks.