Research On Safety Helmet Wearing Detection Based On Convolutional Neural Network

The accident rate of housing and municipal engineering has been on the rise in recent years,and the safety of construction personnel is the focus of the country’s attention.Regulatory regulations and the cultivation of safety awareness are not the key to solving the problem.Only by implementing and improving the detection mechanism for violations of regulations can this problem be solved fundamentally.Wearing a helmet is the most basic measure to protect the safety of personnel in infrastructure construction projects,and the detection mechanism for whether the helmet is worn or not is also a safety line of defense for the staff.With the rapid development of deep learning,practical applications in the field of target detection can be seen everywhere in daily life.This article will use target detection technology to detect the wearing of helmets.The essence of the target detection problem lies in classification and positioning: to identify whether a picture contains a target,and to mark the area where the target is located.The traditional target detection process is to extract the area,manually select the features,and use the classifier to classify the three steps.There are two strategies for target detection in the deep learning stage.One is to directly regress position and category,and the other is to first generate candidate regions and extract features,and then put them into the classifier for position correction.The former has faster recognition speed but lower accuracy;the latter has better accuracy but slower speed.Based on actual needs,this article takes the YoloV3 model as the main body for experimentation.This paper uses the YoloV3 algorithm as the basis to detect the helmet.Aiming at the hardware computing performance in the actual deployment that cannot meet the actual model application,the channel screening method is used to screen the batch normalization layer after the convolution layer in the entire YoloV3,and the trained The model is sparsely processed,and the trained model is pruned according to three different levels of pruning schemes.Finally,the method of knowledge distillation is used to improve the accuracy of the model to achieve the purpose of actual use.In addition,this paper is based on the YoloV3 algorithm to improve,using the k-means clustering algorithm to re-cluster the anchor frame(anchor),the DIOU loss function to optimize,the densely connected network replaces the original residual structure,and the CBAM module is embedded in the backbone network.The original YoloV3 model was improved,and the effect of each improved module was verified through ablation experiments,and the overall improvement effect of the network was verified through comparison experiments.The experimental results show that the channel pruning strategy in this paper has a more significant effect on compressing the model volume,and the layer pruning strategy improves the forward reasoning speed more obviously,and the detection effect is stable on a simple data set.The safety helmet detection method based on the improved YoloV3 proposed in this paper can significantly improve the safety helmet detection effect.