Chinese Traffic Sign Detection Research Base On Convolutional Neural Networks

Traffic signs contain abundant instruction information for driving and play a very important role in the transportation system.Traffic sign detection is one of the important tasks of traffic sign recognition system.There are various factors which would affect the detection results,such as illumination condition,occlusion.What's more,it is hard to widely apply to practical production due to the lack of real-time.According to the characteristics of traffic signs,traditional traffic sign detection algorithm gets the region proposals by hand-craft features and classifies them to attain the positions and labels of traffic signs by appropriate classifiers.However,the hand-craft features have certain limitations,and the traditional methods have large amount of calculation and tedious process.So generalization ability and real-time of the methods is not high.The object detection algorithm based on deep convolutional neural network can learn the object features from a large number of data,achieve multi-class object detection and get fast detection speed as well.The paper analyzes the advantages and disadvantages of some deep object detection algorithms and further study of the basic theory of convolutional neural network.Furthermore,the paper improves network model based on the existing convolution neural network and proposes a detection method suitable for Chinese traffic signs.Firstly,in order to better study and evaluate the performance of algorithms,the paper makes the Chinese traffic signs dataset,CCTSDB(CSUST Chinese Traffic Sign Detection Benchmark).The data augmentation is adopted to expand the dataset and enrich the diversity of samples,which includes the changes of scale,illumination condition and noise.The whole dataset can be downloaded at https://github.com/csust7zhangjm/CCTSDB.Secondly,the paper proposes a real-time multi-class traffic sign detection method which combines the characteristics of traffic signs to improve the network structure based on YOLOv2,a deep object detection algorithm.The modified network makes full use of multiple 1×1 convolution layer in intermediate and simplifies the structure of top layers.In addition,the modified network divides the output feature maps of the top convolutional layer into finer patch,which could attain fine-grained feature maps to detect small size of traffic signs.Contrast experiments both on CCTSDB and GTSDB(German Traffic Sign Detection Benchmark)indicate that the modified model has the fastest detection speed and higher precision.Finally,the paper summarizes the development of convolutional neural network structure and design strategies to improve the real-time Chinese traffic sign detection method for increasing the detection precision furtherly.To make the most of the minutiae features of traffic signs,the paper proposes a new network model based on Fire module,which integrates the Fire module into the bottom layers of network structure to enhance the network ability to learn shallow features.Contrast experiments on CCTSDB shows the new network model improves the precision and maintains a fast detection speed.