Research On The Detection Method Of Helmet Wearing In Construction Site Based On Deep Learning

While the construction industry empowers my country’s economic development,the problem of construction safety accidents has become increasingly prominent,and about 90% of the safety accidents are closely related to the unsafe behavior of operators.Among the unsafe behaviors on the construction site,the risk of safety accidents caused by not wearing safety helmets according to the regulations is the most serious.It is of great significance to study the detection method of safety helmets worn by construction site workers to ensure the safety of construction industry.This paper mainly focuses on the safety helmet detection of construction site workers,and conducts research on the construction site safety helmet wearing detection method based on deep learning.The main research contents are as follows:First of all,in view of the single scene problem of the existing hard hat wearing detection dataset,the construction site picture data is collected,and combined with the public hard hat detection data set into a mixed data set,and the hard hat wearing detection sample library is obtained through the data enhancement method.And according to the characteristics of the sample data in the data set,the Cascade R-CNN algorithm is determined to be the basis of the helmet wearing detection algorithm.Secondly,in view of the poor adaptability to complex scenes in the Cascade RCNN hard hat wearing detection results,a hard hat wearing detection algorithm based on the improved Cascade R-CNN is proposed.For the problem that the background similar to the foreground is wrongly detected,the Swin Transformer feature extraction network is used to replace the original backbone network Res Net-101;for the problem that small targets and occluded targets are difficult to detect,the ROI Align method is used to improve the pooling of regions of interest;For the phenomenon of redundant or missing target frames,the Gaussian weighted Soft-NMS algorithm is used;in order to enrich the diversity of the data set,the Mosaic data enhancement method is introduced during training;Optimal Threshold Matching.Finally,the improved model is trained and tested on the constructed "My-SHWD" data set.The experimental results show that the m AP of the improved model reaches 92.66%,which has a good detection effect and strong robustness.Finally,based on the improved Cascade R-CNN helmet wearing detection algorithm,combined with the actual demand analysis,a helmet wearing detection system is designed.The system can perform picture test,video test and camera test.The reliability and stability of the system are verified through the test,and the improved algorithm has good practicability.