Research On Hardhat Wearing Detection Based On Deep Learning

Due to the labor-intensive,high working intensity,complex working environment and low automation,casualties frequently emerge in the construction industry.As a preventive measure,the correct use of safety hardhats can effectively avoid traumatic brain injury from accidents and protect the lives of workers.However,many workers often do not wear safety hardhats during working due to carelessness.Therefore,automatic hardhat wearing detection based on computer vision is helpful to strengthen safety management of construction sites and to protect workers’ safety.This thesis analyzes the existing researches on hardhat wearing detection in details.To solve the shortcomings of the existing researches in detection accuracy and generalization,two novel single-step hardhat wearing detection networks are proposed,which are both anchor-free deep learning networks.1.The existing hardhat wearing detection models are mainly anchor-based networks,which are poor in generalization.This thesis proposes an anchor-free deep learning framework for hardhat wearing detection,which is based on Centripetal Net.A split channel attention convolution block is proposed to combine high-efficient convolutions and channel attention,which has the advantages of the improvement of the ability of feature extraction and the reduction of the computing resource consumption for the backbone network.The proposed framework can effectively improve the overall accuracy of the detection model and reduces the memory occupation.2.In order to achieve the balance between detection accuracy and detection speed,a novel framework named Center Attention-Centripetal Net（CA-Centripetal Net）is proposed.It is motivated by strengthening the ability of the detection network to comprehensively utilize internal features and marginal features of objects without additional detection cost.The network combines a light-weight backbone network,a vertical-horizontal corner pooling,and a bounding constrained center attention module only used in the training stage.Experiments indicate that the CA-Centripetal Net has better performance than the existing single-step methods,which has the advantages of real time and small model size.