Research On Road Sign Detection Algorithm Based On Deep Network

With the rapid development of the economy,the total quantity of vehicle in our country has increased greatly in recent years,which is followed by traffic congestion and frequent accident.Therefore,how to solve the management problem in the high load traffic demand under the limited human resources is the top priority.Under this demand,many scholars have combined the development of computer hardware with deep learning to develop solutions such as vehicle-mounted assistance systems and unmanned driving.As an important support for intelligent navigation and unmanned driving,road sign detection is one of the current research hotspots.Road sign detection is mainly to accurately identify the position and category information of road signs in different pictures,and since road signs take up a small proportion in the pictures,it is also small target detection.With the development of road target detection algorithms,from traditional algorithms to deep learning algorithms and from two-stage to single-stage road sign detection models,accuracy and recognition efficiency have been balanced,but there are still defects when it comes to a certain problem.This thesis explores and summarizes the road sign monitoring model at the present stage in a multi-directional way,and combines other methods to improve and innovate.The main work is as follows:Due to the problems of accuracy and missing detection in the road sign detection model,a road sign detection model Yolo4-sc based on attention mechanism was proposed.Based on the single-stage Yolov4 target detection model,CBAM attention mechanism was added into the feature extraction network to apply different weights to the spatial dimension and channel dimension.To improve the attention to the feature map,the fusion of scale balance pyramid in the feature fusion network can effectively extract cross-level semantic information,self-adapt the degree of ambiguity of adjacent layers,maintain the scale balance between the feature map,without increasing the amount of calculation while improving the accuracy.A large number of experimental results show that compared with other classical models,the Yolov4-sc model designed in this thesis has a superior road identification effect.Compared with the original model Yolov4,the mAP has been improved by 1.91%,and the accuracy of the motor vehicle lane class and the no-driving class has reached 98.6%and 98.5%.The model is stable and reliable,with high precision and improved network performance.Road sign detection model is mainly applicable to small equipment or embedded equipment because of the limitations of computing power and low accuracy of small equipment.Based on this problem,lightweight network Yolov5s is selected as the improved model,and the idea of Swin Transformer is added to the model to propose the Yolov5s-swin model.First,Swin Transformer module is integrated with convolution.Resunit module in the backbone network of Yolov5s was improved,Cswin module was proposed,and PAFPN in the network was modified to enhance the ability of the model to capture local feature information and improve the accuracy of network detection.Experimental results show that the detection accuracy of the Yolov5s-swin model proposed in this thesis is improved by 2.4%.Compared with existing models,the road sign detection effect is better,and the FPS reaches 128 frames per second,which can meet the real-time requirements of the equipment.