Research On Traffic Sign Detection Technology Based On Lightweight Convolutional Neural Network

Traffic signs are essential to road traffic safety.There are many influences in the actual driving scene,such as the change of light,the deformation of traffic signs,speed.As a result,people’s eyes may miss or mistakenly identify traffic signs,resulting in the wrong judgment of the road conditions ahead,causing traffic accidents,causing the loss of personal property and vehicles,and even threatening the safety of life.As a crucial part of the advanced driving assistant system,real-time and accurate traffic sign detection technology can help drivers ensure driving safety and avoid danger.It has essential applications in traffic safety and automatic driving.With the rapid development of deep learning and information technology,new traffic sign detection methods boost.The traffic sign detection algorithm based on convolutional neural networks has achieved significant breakthroughs,causing an unprecedented research boom.The deep convolutional neural networks have huge internal parameters and multi-layer nonlinear mapping,combined with the training of large data sets,enable deep models to have high-quality feature extraction and expression capabilities.However,due to the massive number of parameters and calculations,these models often have high requirements for hardware computing capabilities and require high-performance servers.It is challenging to deploy a deep model on a hardware platform with limited resources in vehicles.So,the deep detection models are rigid to meet the needs of practical applications.This paper conducts in-depth research on the problems to be solved and improved in the lightweight traffic sign detection model.Furthermore,the influence of convolution methods,network structure,loss function,and other factors on the detection performance are explored in detail.The research content mainly includes the following three aspects:（1）For the feature extraction network of the lightweight traffic sign detection model,there is a problem of severe information loss leading to insufficient feature extraction capabilities.Based on Dense Net,an improved lightweight feature extraction network is proposed to ensure the network while having a lightweight network structure with an excellent ability to express features.（2）The lightweight traffic sign detection model’s insufficient ability to learn features of small objects leads to a low regression rate of detecting small objects.We proposed an improved lightweight version based on the Det Net network,which combines increased receptive fields,feature fusion,and optimized loss function.In this way,the model’s ability to learn features of small objects is improved,and the small object detection performance of the model is improved.（3）Because of the insufficient accuracy of the Mobile Netv2-SSD detection model in the detection of traffic signs,it is improved by multi-scale pixel feature fusion and efficient channel attention mechanism,which increases the number of parameters and hardly affects the speed.Detection accuracy,build high-precision real-time traffic detection algorithm.