Research On The Algorithm Of Electric Bicycle Helmet Wearing Detection Based On Deep Learning

Electric bicycles are popular for their flexibility,convenience and cheapness.However,at the same time,various traffic violations related to electric bicycles have emerged,and the problem of cyclists not wearing helmets is particularly prominent.To address this problem,there is an urgent need to develop a target recognition algorithm that can efficiently and accurately detect whether a cyclist is wearing a helmet or not.In recent years,deep learning has become a research hotspot because of its excellent feature portrayal ability.Combining deep learning with target recognition can effectively reduce the labor cost.In this paper,we take the cyclist wearing helmet as the research object,and aim to effectively improve the detection accuracy of electric bicycle helmet wearing.A two-stage helmet detection method based on deep learning is proposed for e-bike riders whose head area occupies a small area in the actual traffic monitoring video image and is easily obscured by each other,which can easily cause inaccurate detection and missed detection.The main research content and innovation points of this thesis are as follows:(1)In the first stage,YOLOv4-based e-bike rider detection algorithm is proposed.First,based on the YOLOv4 model,Ghost Net is adopted as the backbone network to reduce the model parameters and improve the detection speed;subsequently,the Efficient Channel Attention(ECA)module is introduced in front of the three-layer prediction head to make the model focus on e-bike cyclist features more effectively and improve the detection accuracy;finally,the neck network part of the Finally,the Spatial Pyramid Pooling(SPP)module is replaced with the Receptive Field Block(RFB)module in the neck part of the network to obtain a larger sensory field to improve the feature learning capability of the lightweight network.The experimental results show that the improved YOLOv4 model has improved in precision and speed compared with the conventional YOLOv4 model,in which the mean average precision(m AP)is increased by 1.39% and the number of frames transmitted per second is increased by about 14 frames.(2)In the second stage,Center Net-based helmet wearing detection algorithm for cyclists is proposed.Taking the cyclist images detected in the previous stage as input,firstly,based on the Center Net model,the shallow features are fully utilized by adding the Feature Pyramid Network(FPN)structure and fully fused with the deep features to improve the detection accuracy;subsequently,the Distance-Io U(DIo U)loss function is employed)loss function is employed to replace the target size loss function to improve the detection accuracy of a single prediction frame.The experimental results show that compared with the conventional Center Net model,the improved Center Net model has a significant improvement in accuracy with a 2.57% increase in m AP to 97.74%.