The Research On Small Target Traffic Sign Detection Technology In Natural Scene

With the traffic pressure of urban roads and people’s pursuit of high-tech life,intelligent driving system has always been an important research content.As one of the important branches in this field,the research on traffic sign detection technology is equally important.Nowadays,traffic sign detection mainly faces the challenges of small detection object,complex detection scene,poor real-time detection and so on.For these problems,this thesis chooses convolutional neural network as a tool to study.In order to improve the learning ability of features,this thesis takes the basic detection network as the framework,strengthens the depth of feature extraction,and improves the utilization of features through different scale feature fusion and multilevel prediction.For the problem of poor real-time detection,this thesis uses a single-stage target detection method.The main contents of this thesis are as follows:(1)For the problems of small target and natural environment interference in road traffic sign detection,an improved multi-scale feature fusion detection network based on FCOS algorithm,FCOS-t network,is proposed.The FCOS algorithm is selected as the framework,the attention module CBAM is introduced into the backbone feature extraction network,and the Swish activation function with better model effect is introduced.Lightweight multi-scale feature fusion in feature enhancement network.According to the scene characteristics of traffic signs,the fuzzy sample processing of the original FCOS algorithm is effectively deleted,and the redundancy of the detection network is reduced.The experimental results show that the FCOS-t network is effective by comparing and analyzing the experimental data with other traffic sign detection algorithms on the Tsinghua Tencent100 K datasets.(2)In order to realize the industrial application of traffic sign detection in intelligent driving system,a fast response lightweight network model,m-YOLOX network,is proposed.The network takes the YOLOX algorithm as the framework,adopts the lighter Mobile Net V3 network,replaces the cspparknet feature layer in the original algorithm with the effective feature layer of Mobile Net V3,and uses the deep separable convolution to replace the ordinary convolution in YOLOX,so as to reduce the parameters of the network model.Finally,the algorithm of m-YOLOX network before and after the improvement is compared on the Tsinghua Tencent 100 K datasets,which proves the feasibility and effectiveness of the network.(3)According to the application scenarios of FCOS-t network model and m-YOLOX network model,FCOS-t model and m-YOLOX model are imported respectively,and a convenient traffic sign detection platform system is designed.The framework and operation process of the system are introduced in detail,and the realization of the image detection function and real-time video detection function of the operating platform system are analyzed in detail.Finally,the different functions of the detection platform are tested,and the operation results are displayed through the visual page to prove the feasibility of the system operation.