Research On Detection Algorithm Of Road Safety Helmet Based On Deep Learning

Because of its convenience and affordability,electric vehicles are the preferred way for most people to travel on a daily basis.In electric vehicle traffic accidents,the rate of craniocerebral injury and mortality among cyclists who do not wear safety helmets is much higher than that of cyclists who wear helmets.In 2020,in order to ensure the personal safety of electric vehicle cyclists,the Ministry of Public Security launched the "One People One Helmet" operation,arranging traffic control personnel to inspect the helmets of electric vehicle cyclists at important intersections.However,due to the large road supervision area and the high traffic flow during commuting peaks,the efficiency of manual supervision is extremely low.Therefore,this thesis conducts research on the detection of road safety helmets,aiming to realize intelligent detection of whether electric vehicle cyclists wear helmets.It is of great significance to improve the efficiency of "One People One Helmet" inspectors,reduce the workload of on-duty personnel,move towards digital traffic management,and maintain road traffic safety.The main research contents are as follows:(1)In this thesis,a method for constructing road safety helmet datasets based on image enhancement is proposed to solve the problem of missing datasets.First,the actual road application scenarios are analyzed,and the images of the dataset are collected by means of web crawlers and flyovers,and the labels are determined according to the detection tasks,and the original data source set is marked;then,in order to increase the number and background richness of the dataset samples and supplement the severe weather samples,the image enhancement operations such as random rotation,brightness adjustment and weather simulation are performed on the dataset to complete the construction of the road safety helmet dataset.Comparative experiments show that it is necessary and effective to use the image enhancement method to expand the original data source set of road safety helmets in this thesis.(2)In this thesis,an improved method based on multi-scale feature fusion and ECA attention mechanism is studied to solve the problems of poor recognition effect and low detection accuracy of YOLOv5 s algorithm in road safety helmet detection tasks.First,analyze the detection requirements of road safety helmets,and select the YOLOv5 s algorithm as the benchmark network for road safety helmet detection;then,for the problem that the model is prone to miss detection of small targets,improve the multi-scale feature fusion module,and add a minimum object detection layer in the Neck part to improve the detection of very small objects.The detection capability of the object and the addition of two horizontal jump links make the feature map of the small and medium-sized targets to be detected fuse the original feature map of the same size extracted by the backbone network,fuse more feature information,and improve the detection effect of small and medium-sized objects;secondly,in order to make the model focus on the key features of the target,strengthen the feature fusion effect,and introduce the ECA attention mechanism in the Neck part of the model;Finally,in view of the problems that the above-mentioned improvement methods increase the complexity of the model and the detection time,the model compression method based on GSConv is studied,which effectively reduces the complexity of the model and speeds up the detection speed.The ablation experiment verifies that the three improvements proposed in this thesis are all effective,and the combined use effect is better;the comparative experiment shows that the optimized model detection effect in this thesis is outstanding and has good real-time performance.(3)In this thesis,a road safety helmet online detection system is built to solve the problems of low efficiency of manual inspection of road safety helmets.Based on the optimized road safety helmet detection model,the system realizes functions such as login verification,picture detection,video detection,data lake visualization and user management.After testing,the system can run stably and realize real-time online detection of the helmet-wearing status of electric vehicle riders in road surveillance videos.This thesis contains 73 figures,10 tables and 86 references.