Helmet Wearing Detection Based On Improved CenterNet With Enhanced Associations

Wearing safety helmets is a safety standard that must be strictly enforced during construction processes such as electricity and buildings.In recent years,the number and scale of construction sites have been increasing due to the strong investment of the country in infrastructure construction,and efficient safety inspection methods have become an urgent need.Compared to traditional time-consuming and laborious manual inspection,machine vision-based methods have the characteristics of high automation and easy expansion,making them one of the current research hotspots.The challenges faced by current automated helmet detection methods include the fact that helmet targets are often small,dense,and partially occluded,making it difficult to balance the accuracy and real-time performance of the method.To address these issues,this article proposes a highly accurate,real-time,and lightweight automated helmet detection method based on the anchor free frame Center Net model,which is embodied in the design of the backbone network,the improvement of model context and attention,and the model lightweight strategy.The specific research work and innovation points are as follows:Firstly,an Association Fusion(AF)architecture is proposed to solve the problem of information loss during convolutional downsampling in backbone networks.To give full play to the feature of pixel-by-pixel classification in Center Net,a backbone network structure with AF is designed to achieve the fusion of deep and shallow features of the backbone network and compensate for information loss.Specifically,it removes the redundant short connection structure in cross-layer connections,increase the size of convolutional cores after cross-layer connections,and adjusts the number of feature graph channels by reducing the number of convolutional cores,maintaining a good balance between improving accuracy and lightweight models.Secondly,aiming at the problems of weak semantic information in shallow networks and insufficient model context capabilities,a Context and Attention(CA)module is proposed.The CA module superimposes a dual attention mechanism of spatial and channel attention,as well as a spatial pyramid pooling module,which can effectively enhance context extraction performance.By embedding the CA module into the AF architecture,the attention guidance of the encoding layer to the decoding layer is enhanced,improves the model context extraction ability,further improves the performance of the model,and alleviates the weak detection ability in certain special scenarios.Finally,aiming at the actual needs of helmet-wearing detection,an efficient and lightweight application system is constructed by combining the AF backbone network structure with the CA,and using Mobile Net V2 to replace Res Net-18 in the original CenterNet as the encoding network.This method outperforms mainstream comparison methods in terms of accuracy,average reasoning time,and average weight on SHWD public dataset.Experimental results show that the proposed method has both strong real-time performance and high accuracy and is suitable for helmet detection in complex scenes.