The Research Of Traffic Sign Recognition Based On Modified Convolutional Neural Network

With the increase in car ownership and driver numbers,the number of traffic accidents has also remained high.Traffic sign recognition is one of the core technologies of advanced driver assistance systems.High-accuracy and real-time traffic sign recognition can promptly give feedback to the driver according to the vehicle’s road environment and assist drivers in driving decisions,which reduces even avoids traffic accidents.On the other hand,traffic sign recognition is also one of the key technologies for the research and development of driverless vehicles in recent years.Therefore,the study of traffic sign recognition is a topic that has important research value and practical significance.The article gives a detailed summary and analysis of the detection and identification methods of traffic signs at home and abroad.It concludes that recognition method based on convolutional neural network has achieved good accuracy in recent years.A series of pre-processing operations on pictures of German traffic sign recognition database,including image cropping to remove redundant backgrounds,image enhancement and image size normalization with contrast-constrained histogram equalization.Finally it gets a high quality dataset image.After detailed analysis of the network characteristics,hierarchical structure,and hyper-parameters of convolutional neural networks,it introduces a classic convolutional neural network model that has achieved good results in image recognition.Finally,on this basis,the paper proposes an modified convolutional network model.The new modified convolutional neural network model has improved the data normalization method,Dropout layer and convolutional channel number,and the effectiveness of improved solution is verified in actual tests.Comparing and analyzing through experiments,it determines the optimal hyper-parameter of the network model and completes the optimization of preset parameters for a new modified convolutional neural network model.The final modified convolutional neural network model achieved a 99.17% recognition rate on the German traffic sign recognition library set,and the average time per image is only 7 milliseconds,which meets the requirements for accuracy and real-time performance.