A Study On Convolutional Neural Networks For Traffic Sign Recognition

Traffic sign recognition plays an important role in intelligent transportation systems and it is an essential component of driver assistance systems and autonomous vehicles.Recently,great achievements have been made by convolutional neural networks in the field of computer vision.Traffic sign recognition based on convolutional neural networks also attracts people's attention.In the thesis,various network parameters are tested,the working mechanism of the network is studied and a method to simultaneously extract the foreground is proposed.The main works are as follows:First,a study of network structures and network parameters are carried out.LeNet-5 and AlexNet network are compared based on GTSRB traffic sign recognition database,which turns out that LeNet-5 seems a good option.Parameters such as convolution kernel sizes,number of convolution kernels,number of batch samples,and number of full connection layers are optimized based on the LeNet-5 network.Furthermore,to improve the speed of the method,parameter compression is carried out,and one third of parameters in full connection layers are reduced.Second,the working mechanism of neural networks is studied based on two methods.By the visualization of convolution kernels in each layer of the network and a careful study of the neuron responses for different input images,the function of the neurons in each layer is discussed,which is helpful for the network design.Finally,a method to simultaneously obtain the classification and segmentation results is developed.With no extra segmentation annotations,foreground regions of the traffic signs are extracted.Preliminary results are provided to validate the efficiency of the method.Overall,the parameter tuning technology of convolutional neural networks for traffic sign recognition are studied.Based on the study of the network working mechanism,a new method is proposed for simultaneous classification and segmentation of traffic signs.The developed methods are also useful for other object detections based on convolutional neural networks.