Research On Real-time Helmet Wearing Detection Method Based On Optimized YOLO V5

In recent years,the probability of safety accidents in the construction industry remains high,not wearing helmets correctly is the main cause of casualties.Most construction sites are mainly supervised manually for helmet wearing inspection,and this traditional way is costly and inefficient.With the development of deep learning technology,many scholars have carried out research on helmet wearing recognition,most algorithms are not well adapted to the complex environment of the construction site,and is limited to algorithm research,not applied to the actual detection scene,the number of security systems on the market for helmet wearing real-time detection is relatively small.Therefore,this paper designs and implements a safety helmet real-time detection system based on the optimized YOLO v5 algorithm,which can make accurate and fast identification of safety helmet wearing situation at construction sites,which is important to protect workers’ life safety.This paper compares and analyzes the performance of four different networks in the newly introduced YOLO v5 algorithm,selects the more lightweight YOLO v5 s network,and improves the a priori box clustering algorithm based on it,selects the appropriate CIOU＿Loss as the loss function in prediction,optimizes the original YOLO v5 s.formulates the screening principle,cleans and image enhances the open source safety helmet data images,and obtains a total of 70 safety helmet wearing data sets.A total of 7030 helmet wearing datasets were obtained.The Py Torch framework and the optimized YOLO v5 s algorithm are then used for training,and the hyperparameters are continuously adjusted and a suitable optimizer is selected to improve the model performance.The final obtained model has a volume of only 14.4M,a detection speed of 97.7fps,and a detection accuracy of90.1%,and a helmet wearing real-time detection system is built using the model.The results of the work in this paper are as follows.(1)A hybrid data enhancement method applicable to helmet detection is proposed.For different seasons,different weather and different time of day on the construction site detection environment,change the brightness,saturation,color temperature and clarity of the picture to simulate different detection environments,reflecting the complex situation in the real monitoring screen of the construction site.A hybrid data enhancement method is proposed: on the basis of the commonly used linear transformation to process images,data processing is performed on one picture by various methods such as hybrid noise addition,HSV random transformation and random rectangular masking.This method not only expands the number of data sets,but also improves the diversity of helmet data sets in real construction scenes and generalizes the detection capability of the model.(2)K-Means++＿(1-IOU)clustering algorithm is proposed to optimize the prior frame.Real pictures of construction sites are characterized by small,numerous and densely distributed targets to be detected,which require more accurate size and location of the a priori frame prediction.In response to the defect of inaccurate identification of the center position by Anchors produced by the traditional K-Means clustering method,a K-Means++＿(1-IOU)clustering algorithm is proposed that is closer to the characteristics of the sample data: it incorporates the advantages of K-Means++,a derivative version of K-Means,in optimizing the center problem,and uses 1-IOU to replace the original K-Means algorithm in of the Euclidean distance,this method yields Anchors that are more consistent with the distribution of the sample data and improves the recall of the model.(3)A helmet real-time detection system is designed and implemented.The system structure,functions and interface were designed according to the demand of helmet real-time detection,the overall framework of the system was built and the detection process was designed,the trained model was deployed into the system,and the automatic detection of helmet wearing with static pictures,historical videos and real-time camera access was realized.After functional testing and real construction site verification,the system has high detection accuracy and anti-interference effect.