| In recent years,the safety accidents in the construction industry,about 60% of the casualties were caused by failure to wear safety helmets in China.Wearing a safety helmet is of vital importance for guaranteeing the life and production safety in various engineering environments,and requiring workers to wear safety helmets has also become a necessary measure.However,the existing manual supervision method of wearing safety helmets lags far behind the current information-based and intelligent production mode required by Industry 4.0.How to realize the intelligent detection of the wearing safety helmet has become a hot issue waiting to be solved.Based on the advanced deep learning target detection technology,this paper proposes a intelligent detection method for safety helmet wearing in the field of computer graphics through improving related algorithms.Through a comparative analysis of recent deep learning target detection models,this paper selects the YOLOv3 algorithm as the main research object,based on which the network structure,loss function,and related functional modules of the YOLOv3 algorithm are improved.In terms of the problem of the small size of the helmet image in the existing data sets and actual acquisition,this paper introduces the spatial pyramid poolin structure(SPP)and the shallow feature fusion layer that optimizes the overall performance of the YOLOv3 algorithm.Considering the disadvantages of the regression loss function,the GiOU idea is introduced to improve the original loss function and optimize the regression effect of the network model.Furthermore,this paper analyzes the shortcomings of the NMS algorithm and K-means algorithm in YOLOv3,and replaces it with Soft-NMS and K-means++ algorithms.In terms of the problem of the small size of the helmet image in the existing data sets and actual acquisition,this paper introduces the pyramid pooling structure and the shallow feature fusion layer to optimize the overall performance of the YOLOv3 algorithm.Considering the disadvantages of the regression loss function,the GiOU idea is introduced to improve the original loss function and optimize the regression effect of the network model frame.Furthermore,this paper analyzes the shortcomings of the NMS algorithm and K-means algorithm in YOLOv3,and replaces it with Soft-NMS and K-means ++ algorithms.Based on the improved YOLOv3-SPM-G network,this paper conducted experiments and analysis,and achieved a relatively ideal detection effect.The average accuracy precision(m AP)was increased from 79.7% to 88.6% of the original network,and the detection speed reached 43.22 frame per second.The YOLOv3-SPM-G network model can achieve better detection results under shelter,small sized targets,similar interference and dense targets,and meet the needs of real-time video surveillance,making it have broad application prospects.Finally,in order to intuitively show the effectiveness of the detection algorithm,this paper develops a safety helmet detection interface based on PyQt5 under the Linux environment. |
