Research On The Speed Limit Signs Recognition Based On The Convolutional Neural Networks And Self-warning System

Speed limit sign is one of the common and important signs of prohibition.It provides the limit speed for drivers to guide them drive safely.But in most time,drivers would ignore the speed limit signs.For example,using mobile phone in the car,talking to passengers or the speed limit signs setting in sight of the blind would cause the vehicle speed too fast.Then it would cause traffic violations and traffic accidents.Therefore,the development of speed limit sign recognition and early warning system can effectively reduce the occurrence of traffic accidents and ensure the safety of life and property.Based on the Convolutional Neural Networks,the speed limit signs recognition and self-warning system in the vehicle is developed.The details are as following:This paper analyzes the current situation of the research on traffic sign recognition,deep learning and related applications at home and abroad,and summarizes the characteristics of speed limit signs,such as the shape,color,setting and the location features.Then,the paper analyzes the requirements of the potential users and the technical difficulties of the system development.The operation method and the operation interface of the system are introduced.By analyzing the current situation,the corresponding hardware equipment configuration is selected.And then through the analysis of their own characteristics and advantages,select Open CV as a system development database,select the Hierarchical Data Format as a programming process for image processing data format.Images preprocess and speed limit signs position.Firstly,according to the running state of the vehicle and the driver’s visual characteristics,the processing frequency of the program is analyzed.According to the speed limit signs distribution characteristics,traffic signs rough positioning.Then,the color segmentation is carried out based on the HSV spatial color model,then the morphological filtering is carried out,and the image is preprocessed.The circular mark is precisely positioned according to the method of length and width detection and Huff circle transformation.And then using the speed limit signs` gray scale and digital distribution characteristics to exclude the pseudo-target,so as to accurately locate the speed limit signs position.The image of speed limit signs is identified based on convolution neural networks.Firstly,image enhancement and normalization are performed on the captured image.Also,it is not necessary to tilt correction the speed limit signs by analyzing the large number of images.Then,the convolution neural networks model is constructed,which is the Le Net-5 model.It include two convolution layers,whose convolution kernel is 5×5,two pooling layers,which using max-pooling,the dropout layer and the all-connected layer.The activation function is the Re LU function.Softmax classifier is used for classification.Then we use the GTSRB and the partial images collected by the real test as the data set.Lastly,the system is trained and tested based on the Caffe depth learning framework.The speed limit signs real-time recognition and self-warning system is systematically developed and the performance is tested.First,the development process of the system is introduced.Secondly,based on the Caffe depth learning framework,the system is trained and tested.Then the results are analyzed and summarized.The relationship between the training rate,the test accuracy rate,the loss function and the number of iterations is analyzed.During the training process,the system present a good convergence,and the test accuracy reach to 97% during the test.Finally,the performance of the system is tested.After testing,it was found that the system had a strong compatibility with different android devices,and the performance of the system was verified by road test.In the different speed state,the accuracy rate of the speed limit signs real-time recognition and self-warning system can reach more than 85%,which can up to 94%.And through the test process,found that it has a strong real-time,system stability and identification robustness.