Research On Detection And Application Of Safety Helmet Wearing Based On Deep Learning

With the existing Deep Learning algorithms is applied to the detection of wearing a helmet in a complex construction site,the detection speed,accuracy and robustness are poor.At the same time,the existing algorithm haven’t add a function of alarm.To prevent serious injury to workers when accidents occur,based on Deep Learning,an application system is designed and built for wearing helmets in this paper.First,the Two Stage method and the One Stage method in deep learning are studied.According to the characteristics of the detection object,the model is improved to enhance the detection accuracy and robustness.After comparing the two methods,the model with excellent performance is selected for security in the helmet wearing detection application system.An alarm module is added to the system when the worker is detected not wearing a helmet,an singal is generated.At the same time,the signal is transmitted to the alarm module through network communication.Receiving the signal,an audible and visual alarm is generated in the alarm module.1.First,the research on Faster-RCNN is developed in the Two Stage method.Then according to the characteristics of the research objects in this article,the key points of the algorithm is relevant improved,specifically: using Res Net101 instead of VGG16 as the feature extraction network of Faster-RCNN;adjusting the size and aspect ratio of the RPN preset anchor;adding OHEM in the training process,optimizing the generation of candidate regions;using ROI Align instead of the ROI Pooling layer to adjust the size of the candidate frame.Finally,it is verified on the test set.The experimental results show that the m AP of the model is imploved to 83.24%,which is 11.43% higher than before,and the accuracy of the model has been improved.2.First,the research on YOLOv3 is developed in the One Stage method.Then according to the characteristics of the research objects in this article,the key points of the algorithm is relevant improved,specifically: adjusting the preset anchor box size,according to the anchor box size modifying the output layer scale,and then adding a Residual layer to the output layer network;using Focal Loss replaces Cross Entropy as the classification loss to solve the imbalance of positive and negative samples;using Soft NMS instead of NMS.Finally,it is verified on the test set.The experimental results show that the m AP of the improved model has increased by 4.84%,and the accuracy of the model has been improved.3.Compared two models,the results show that the detection speed and accuracy of the improved YOLOv3 are higher than the improved Faster-RCNN.The experimental results show that removing the model loading time,the improved YOLOv3 only takes 0.07 S to detect a picture,and the m AP reaches 91.94%,which meets the actual needs of helmet wearing detection in terms of detection speed and accuracy.4.A safety helmet wearing detection application system is designed and built,the system includes: video acquisition module,safety helmet wearing detection module,network communication module and alarm module.The feasibility of the system is verified through the laboratory simulation scene test.When the detection module detects that the person is not wearing a helmet in the video,it generates an alarm signal,transmits the signal to the alarm module remotely through the network communication module,and the alarm module receives it.After then,a voice alarm which can remind workers to wear the helmet properly is issued,and the functional requirements of the helmet wearing detection are realized as a whole.