Research On Helmet Wearing Detection Of Electric Bicycle Drivers Based On Deep Learning

At present,electric bicycles have become a very important part of urban traffic.With the increase of the number of electric bicycles,traffic accidents related to electric bicycles also increase,causing certain personal and property losses to electric bicycle drivers.Wearing safety helmets can effectively protect the life safety of e-bike drivers and reduce the casualty rate after accidents.Relevant government departments have issued corresponding documents and policies to manage e-bike driving behaviors,among which wearing helmets is an important point.At present,the helmet wearing inspection of EV drivers is mainly judged by the manual observation of traffic police,which is timeconsuming and laborious,and has the disadvantages of low efficiency.This paper studies helmet wearing detection of electric vehicle drivers based on deep learning.First of all,in view of the fact that the helmet data set of EV drivers is not disclosed at present,this paper makes an in-depth analysis of the possible problems in the selfcollected data set.Then,the data set was collected through Internet download and manual shooting.In order to meet the diversity of the target and background of the data set,various factors such as different angles,different light intensity and different locations were considered in the collection process,and geometric transformation,brightness transformation,noise addition and other methods were used to realize data augmentation and expansion of the data set,further improving the data diversity.Finally,different data set partitioning methods were compared and the original data was divided into training set and test set by setting aside method,and label Img software was used for data annotation to complete the data set construction.After that,the detection accuracy and lightweight of the model are improved.The first step is to introduce Ghost module into multiple positions of the original network model and conduct experimental comparative analysis.The experimental results show that adding Ghost module into the backbone network has the best effect.It can reduce the number of network model parameters and floating point computation without decreasing the detection effect.In the second step,SENet,CBAM,GAM and CA attention mechanism modules were respectively introduced into the neck network of the original network model,and the experimental comparison and analysis were carried out.The comparative experimental results showed that the CBAM attention mechanism had the best effect in this data set,which could improve the accuracy of the model’s detection of the target,and the number of model parameters was not increased too much.Finally,Ghost module and CBAM attention mechanism are introduced in the above way,and an improved YOLOv5 network model is proposed in this paper for experimental analysis.Before the experiment,K-means method was used to re-cluster Anchor,and the loss function was modified to EIoU.The improved model is compared with the original model.Compared with the original model,the mAP of the improved model is increased by 2.7%,the number of parameters is decreased by about 15%,and the floating-point arithmetic is decreased by about 22%.In other words,there are improvements in the above aspects.Compared with SSD and Faster R-CNN,target detection algorithms also have corresponding advantages.