Research On Detection Methods Of Construction Workers’ Safety Helmets Based On Deep Learning

Safety helmet is an important protective equipment for delaying and relieving the impact of external force on the head,and it is required to wear it correctly in various job sites with safety protection requirements.In recent years,among the safety accidents in major construction sites,construction safety accidents caused by construction workers not wearing safety helmets correctly account for a large proportion.Therefore,it is of great practical significance to prevent and reduce the occurrence of building safety accidents by automatically detecting whether the safety helmet is correctly worn by the personnel entering the job site(hereinafter referred to as "safety helmet detection").Through on-the-spot investigation of workers’ wearing safety helmets in construction engineering operation sites,and after in-depth study and analysis of safety helmet detection methods,the difficulties of research topics are summarized as follows: 1.It is difficult to obtain data sets based on construction engineering scenes;2.The background of the construction work site is complex,and the target detection is difficult;3.Safety helmet detection belongs to small target detection,which requires high accuracy of algorithm.4.Safety helmet detection requires real-time detection,which requires high speed of algorithm.To solve these difficulties,the main research work is as follows:(1)Construction of safety helmet detection data set for construction workers.Data set is the basis of studying safety helmet detection.However,there is no baseline data set of safety helmet based on construction engineering.Therefore,a set of safety helmet detection data set based on the construction scene is sorted out by collecting data from the construction site.(2)Analyze and compare the commonly used target detection algorithms.According to the special scene and detection object of helmet detection,aiming at the problem that You Only Look Once v3(YOLOv3)algorithm is easy to miss small targets and adjacent targets,the multi-anchor mechanism and soft-NMS are introduced into the detection process.Aiming at the problem that the weight distribution of loss function in YOLOv3 is easy to cause error detection,a new weight distribution method of loss function is proposed.The average precision of the improved YOLOv3 algorithm is improved to 88.6%.(3)A small target detection algorithm in complex background is proposed.To solve the problem that Single Shot Multi Box Detector(SSD)algorithm is difficult to detect small targets,a "level" feature extraction method is constructed by combining Conv LSTM algorithm with the feature pyramid structure of SSD algorithm,which can better extract the feature information of small targets and further enhance the ability of judging features.In view of the complex background of the construction work site,attention mechanism is introduced to make the safety helmet inspection process pay more attention to the target features and ignore the background features.In addition,to solve the problem that the training model falls into local optimum due to the imbalance between positive and negative samples,the original loss function of SSD is replaced by Focal loss function.Comparing the improved SSD algorithm with other algorithms,the experimental results show that the average precision of the improved SSD algorithm is improved to 94.7%,which has a good detection effect.(4)Using Tensor RT to accelerate the process of model reasoning,a safety helmet detection system is built,which is used to detect the wearing condition of safety helmet of workers in real time.After testing,the system has achieved the expected goal.