Research On Domestic Traffic Sign Detection And Recognition Algorithm Based On YOLOv3

With the continuous development of science and technology,my country’s per capita GDP continues to increase,which has led to a substantial increase in the number of motor vehicles owned by Chinese citizens.However,while the number of motor vehicles is increasing,the traffic accident rate is also increasing year by year,and traffic safety has also increasingly become a concern of people.In 2018,the Ministry of Public Security proposed to vigorously develop intelligent police and data police.At the same time,the concept of intelligent transportation has also been put forward.Now intelligent transportation system has been extensively used in much more fields like equally national transportation.Traffic sign detection and recognition system play the part of intelligent transportation.In recent years,in the wake of the uninterrupted enhance of CPU and GPU computing capability,it is immediate significance to develop traffic sign recognition algorithm research by utilizing existing public security equipment resources.The convention traffic sign gauging ways has many difficulties in solving the problem of traffic sign recognition in different complex environments.The traditional traffic sign detection algorithm does not have the network structure of lightweight and high recognition accuracy.Therefore,this paper studies the traffic sign detection task based on YOLOv3 framework.However,there are still some difficulties in using YOLOv3 to detect traffic signs in real time: YOLOv3 feature extraction network has a large number of parameters and a long forward reasoning time,which is not conducive to real-time detection;In YOLOv3 network structure,the original FPN network does not make full use of the features of all levels and has poor recognition effect for small objects.Due to the interference of various adverse factors,the quality of some images is poor,and most of the traffic signs are relatively small,which produce enormous throw down the gauntlet to the task of traffic sign detection.Therefore,this paper introduces the current novel detection algorithm method into the detection and recognition of traffic signs,and carries out experiments according to the steps of traffic signs detection in the actual environment.It is seeing that the aforementioned problems,the article has a sea of dedication as follows:1.MobileNetv2 is introduced into YOLOv3 as a foundation feature collection network,and an inspection model based on MN-YOLOv3 is formed.Two down-connections are introduced into MN-YOLOv3 backbone network for feature fusion,so that the fusion can greatly reduce the calculation amount of algorithm parameters.The detection speed of the algorithm is accelerated and the feature fusion between layers is more sufficient.According to the characteristics of the target shape of traffic signs,the k-means algorithm is used to calculate four kinds of aiming frames,and the distance intersection and merge ratio loss function is used to regress the aiming frame.so as to speed up the convergence speed of the network.2.In order to better detect and identify small targets,this paper proposes to create a new prediction branch C4 on the basis of the original network structure.Finally,unified information fusion is carried out on the feature layers C1,C2,C3 and C4 to strengthen the information connection among the layers,which is instrumental in making much of the information of each layer.In order to better fuse the high-level semantic information and the low-level fine-grained information together,this paper introduces the adaptive feature fusion method into MN-YOLOv3 model.Passing through the experimental results demonstrate that the mended algorithm is faster than others and the property is better than others,and the average accuracy can reach 96.20% on the China Traffic signs data set of Changsha University of Science and Technology.By comparison,the former detection algorithms which the improved algorithm has preferable performance in this paper.