Safety Helmet Detection Algorithm Based On Deep Learning

In recent years,the construction manufacturing industry has caused a series of security incidents due to a series of reasons such as dense personnel,complex environment,and weak automated testing capabilities.As a common protective equipment in production and life,safety helmets can protect the life safety of workers in construction scenarios to a large extent,but there are always workers who forget or do not wear safety helmets in actual engineering projects,which will bring hidden dangers to the safety of site operators,and on the other hand,it will also delay the progress of the project.In view of the fact that most of the current construction sites use manual supervision for safety helmet wearing detection,which is timeconsuming,laborious and has poor real-time performance,it has become a research hotspot to think about how to combine deep learning object detection technology with real-time detection of safety helmet wearing.In view of the shortcomings of artificial supervision,this paper proposes two improved safety helmet detection algorithms based on deep learning object detection technology,and the main research work is as follows:(1)Due to the frequent accidents under the current construction site,it is very important for the detection of workers wearing safety helmets,and this paper first describes its research background and significance.Then,the target detection algorithm and safety helmet detection are introduced in detail at home and abroad.The overall architecture and composition of convolutional neural networks are described,and finally the YOLO series of one-stage detection algorithms used in this paper is elaborated in detail,and the iterations of this series of algorithms are compared and analyzed to analyze the current development trend of this series of algorithms.(2)The helmet wearing detection dataset is established,some pictures are selected from the SHWD public dataset,and offline construction scenes are taken using mobile devices such as cameras,and the helmet wearing detection dataset is constructed on this paper.The dataset contains helmet images of targets of various scales,covering actual scenarios such as construction sites and production plants,and the scenarios are more complex and more diverse than the public dataset.(3)A helmet detection algorithm based on improved YOLOv5 is proposed.In view of the series of problems such as low safety helmet wearing detection accuracy and easy to miss detection caused by the complex and changeable environment of the construction site,the single-stage detection algorithm YOLOv5 is selected as the benchmark network(YOLOv5s with the smallest weight file is used).ECA-Net is a super channel attention Module,which allows the model to focus more on the specific category of hard hats;Then,the ASFF module is introduced to make full use of the characteristics of different scales.At the same time,scale detection is increased,the feature extraction ability of the original model in dense and complex scenes is strengthened,and the detection area is expanded,which improves the detection accuracy of small targets such as safety helmets.The experimental results show that the average accuracy of the improved algorithm is improved by 3.compared with the original algorithm on the basis of maintaining a faster detection speed 7%.(4)Aiming at the problems of weak real-time detection algorithm of the current safety helmet detection algorithm,an improved YOLOv7-tiny safety helmet real-time detection algorithm is proposed.Firstly,the EPSANet Block pyramid split attention module is introduced to capture detailed information and make the model more focused on the target features related to training helmets.Secondly,the Tiny-Bi FPN structure with fewer design parameters is used as the feature pyramid structure in the original model feature fusion module,which enhances the multi-scale feature fusion capability of the model and improves the missed detection rate of network security cap detection.Finally,the more advanced localization loss function(SIo U Loss)is used,and the vector angle of the required regression is added to improve the convergence speed of the prediction box during the model training process.The experimental results obtained on the self-made safety helmet wearing detection dataset show that compared with the original algorithm,the m AP value of the improved detection algorithm is increased by 2.89%,and the detection speed is increased by 4.8 frame/s,which realizes more real-time and accurate safety helmet detection requirements.