A Traffic Sign Detection Algorithm Based On Deep Learning

Traffic sign is significant security facilities on the road,which plays an important role in regulating traffic behavior,ensuring the safety of the road and guiding the smooth passage of vehicles and pedestrians,etc.The detection of traffic sign has become a hot spot of the intelligent transportation system.Due to the shortage of traffic sign dataset,the existing detection methods are usually limited to a predefined set of traffic signs.With the current understanding,there is not a algorithm could detect all the 7 classes of Chinese traffic sign.Therefore,a framework of traffic sign detection,data automatic labeling and convolution network model trainning is proposed.The new convolution network model is iteratively generated using this framework.The new model can get a higher detection accuracy.Firstly 726 images of seven main categories of Chinese traffic signs and their subclasses are obtained by manual collecting and labelling.Then brightness transformation,gaussian blur and category equalization processing is used to the original images.Thus,an initial Chinese traffic sign dataset is obtained.A traffic sign detection method based on the convolutional network including the region proposal network?RPN?is presented.The initial dataset is applied to fine-tune three detection models VGG16,VGG_CNN_M1024 and ZF pretained on ImageNet2012.The fine-tuned convolutioanal model can detect 7 categories of Chinese traffic signs.The exprimet analyse the effects of different data preprocessing and different convolution network model.The expreimental results show that the ZF model has the best performance for traffic sign detection.The detection precision of the initial test dataset is 91%.But the precision is reduced to 84%after adding some new test data.This is caused by lack of training data.In order to solve the problem of basic detection algorithm's poor generalization performance,an auto-labeling algorithm is proposed to automatively construct the dataset using the graph cut technology and the above trained ZF model's detection result.Then the auto-labelling method is used to analyze the driving video,automatically collect and label new data to construct a new traffic sign dataset.A new convolution network model is obtained by fine-tuning technology using the new constructed dataset.The experimental results show that the new fine-tuned model has a higher mean average precision than the initial one,it increases 7.5%.Finally the latest model is tested by 33 video sequences recorded by a mobile phone.The model detects 1053 traffic signs of 1057 in the video,has 99.62%detection precision.The results verify the effectiveness of the new model.This data collection-model training process can be iteratively applied until a training model satisfying the requirement of detection precision.