Font Size: a A A

Traffic Sign Recognition Based On Feature Fusion And Deep Convolutional Neural Networks

Posted on:2018-03-03Degree:MasterType:Thesis
Country:ChinaCandidate:Z LiuFull Text:PDF
GTID:2348330536470554Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Vehicle in people's life play an increasingly important role.Safety and smooth operation of driving environment is ideal situation of traffic system.Traffic Sign Recognition is divided into two parts: target location and target recognition of traffic sign,which plays a significant role in Intelligent Transportation System.We treat the traffic sign as research objects,propose traffic sign recognition method based on feature fusion and deep convolutional neural networks.We first introduce domestic and oversea research status of traffic sign recognition.Then,we compare the advantages and disadvantages between methods.Thirdly,we propose target location method based on feature fusion and target recognition method based on deep convolutional neural networks.In the target location issue,we present that extract feature of HOG and LBP.Then,using support vector machine as classifier after serial feature fusion.The method is proved to be feasible in effective location of picture with traffic sign and eliminate interference of picture without traffic sign.Deep convolutional neural networks,which is highly valued since its extremely excellent learning ability and application effect,is a recently propose machine learning method model differs from shallow neural networks.We introduce principles and training methods of Auto Encoder?Sparse Coding?Restricted Bolzmann Machine?Deep Belief Nets and Convolutional Neural Networks.Besides,we emphatically introduce deep convolutional neural networks involve Alex Net and GoogleNet etc.According to the research object and application scene,we propose an deep convolutional neural network model TSR9L-Net for traffic sign recognition and establish appropriate training images database.We take the trade-off of recognition accuracy and speed into consideration then build a 9 layers network with lightweight parameters which have 6 weight layers.We train the sample sets which have 7 kinds of warning signs and15 kinds of prohibition signs,and compare Le Net-5?Alex Net?TSR9L-Net,three model's training effect as the same time.TSR9L-Net has a higher accuracy as well as a fastrecognition speed.In GPU mode,7 kinds of sign's recognition speed up to 29.3ms and accuracy up to 99.09% when run a batch with 40 images;15 kinds of sign's recognition speed up to 32.0ms and accuracy up to 99.29% when run a batch with 40 images.The experiment proves that TSR9L-Net exceed AlexNet in accuracy as well as speed.
Keywords/Search Tags:Traffic Sign Recognition, Feature fusion, HOG, LBP, Deep convolutional neural networks
PDF Full Text Request
Related items