Helmet wear detection based on EfficientNet-Y

Abstract

In response to government policies, "Internet +" has been integrated into the construction site to build a "smart site" ecosystem, and construction departments have begun to carry out visual management of the project. To address the problems of low recognition rate, slow detection speed, high hardware cost and complex construction site background for helmet wearing detection at construction sites, a lightweight model EfficientNet-Y is proposed in order to improve the detection accuracy, enhance the detection speed. The model uses EfficientNet backbone feature extraction network to replace the original. The experimental results demonstrate that the number of parameters of EfficientNet-Y model is reduced by 80% compared with the YOLOv3 model, and the mAP is increased by 1.5% compared with that of EfficientDet. The FPS is improved by 55% compared with YOLOv3 and doubled compared with EfficientDet, while the size of the model is only 1/4 of the volume size of YOLOv3 model. The newly constructed dataset resulted in a significant improvement in mAP for each model, with EfficientNet-Y improving by 7.95%, EfficientDet by 9.58%, and YOLOv3 by 5.01%.