Research On Unconstrained Road Sign Recognition Based On Deep Learning

With the popularity of automobiles,assisted driving technology has gradually received widespread attention.Auxiliary driving technology can obtain current road condition information,remind drivers of the traffic conditions that need attention,and help drivers avoid the risk of accidents.In order to perceive the road situation correctly and quickly,it is necessary to obtain the information of traffic signs.The traditional traffic sign detection algorithm has many limitations,and the effect of traffic sign detection in natural environment is not good.This paper is a deep learning target detection model based on regional recommendation network,which can detect traffic signs under non-restrictive conditions.It is more suitable for assistant driving technology in natural scenes.Firstly,starting with the traffic sign data set,aiming at the lack of a unified traffic sign data set in China,traffic sign images are collected under non-restrictive conditions,and the data set is made and preprocessed and expanded.The data sets produced in this paper are all taken under non-restrictive conditions.The background is complex and changeable,which is more in line with the actual situation.Secondly,the accuracy and speed of model detection are studied in this paper.In order to improve the accuracy of the model,the VGG-16 network with better traffic sign extraction ability is selected as the feature extraction network in this paper by comparing and analyzing the feature extraction network.On this basis,the K-Means++ clustering algorithm is used to cluster the labeled frames in the data set,and the most suitable number and size of anchor frames are selected according to the clustering results,which improves the accuracy of the network.In order to solve the problem that traffic signs with small proportion of area tend to miss detection,this paper proposes a multi-level feature fusion strategy.By combining features to extract the convolution features of the third and fifth layers of the network,it not only retains the network's ability to express the overall information of the image,but also adds the details of the image,effectively solving the problem of missing detection of small traffic signs.Through the above methods,the accuracy of traffic sign detection model is effectively improved.In order to shorten the model detection time,this paper uses deep separable convolution to compress the model and separate the convolution core.Without affecting the accuracy of the model,the parameters of the model are effectively reduced and the detection time of the model is shortened.Through the research in this paper,the accuracy of the model is improved effectively on the basis of not increasing the detection time of the model from two aspects: the accuracy of the model and the detection time.The final detection model can be applied to traffic sign recognition and detection under non-restrictive conditions without being affected by changeable weather,complex background,road sign damage and other factors.Figure [47] table [6] reference [81]...