Research On Traffic Sign Recognition Technology Based On Deep Learning

The road traffic sign recognition system mainly includes two parts: traffic sign detection(TSD)and traffic sign classification(TSC).Improving the detection accuracy and classification efficiency at the same time can make the identification system truly serve for the intelligent transportation system.Complicated detection factors such as lighting conditions,object occlusion,and fading of contour edges will seriously affect the detection effect.How to improve the self-adaptive ability of the color segmentation and how to use the shape information of the sign itself to speed up the detection speed is critical.In addition,the serious uneven distribution of training samples and the inconspicuousness of the test image feature information are also the main problems in the classification of signs based on deep learning networks.Aiming at the detection of traffic sign in complex scenes,a new extraction method of traffic sign's ROI is proposed in this paper which is based on color segmentation with adaptive color threshold and hypothesis test of shape symmetry: 1)an adaptive color threshold is obtained using the cumulative distribution function of the image histogram firstly,and according to it,an approximate maximum and minimum normalization method is designed.It is utilized to suppress the interference of high brightness area and background in image thresholding process;2)the highlight shape feature of thresholding image is transformed into a connected domain feature vector.And a statistical hypothesis testing based on shape symmetry detection algorithm is realized to extract the ROI of traffic sign efficiently.Through testing on the GTSDB(German Traffic Sign Detection Benchmark)dataset,the traffic sign detection rate reaches about 94% with this method.Comparing to other methods,it has higher detection accuracy and faster speed.Besides,it demonstrates better robustness under the complex lighting conditions.At the traffic sign classification stage,this paper starts with the improvement of neural network training set and preprocessing with test sample to improve the classification effect: 1)the original data set is augmented by image processing algorithm to compensate for the defects of insufficient sample in the abnormal scene;2)the test sample is used for many purposes.These types of image enhancement process make it possible to better highlight the characteristics of traffic signs.The classification process uses Tinny-DNN and DLIB respectively for deep learning and testing of test samples.The high performance in classification capabilities of convolutional neural networks enables the classification of traffic signs.The recognition rate of this method on GTSDB is greatly improved,reaching a maximum of 96.5% and the average time of classification is 0.1s,so it achieves an efficient classification.The detection and identification schemes in this paper were tested in German Traffic Sign Detection Benchmark(GTSDB)and Belgium TS for Classification(BTSC)respectively.The experimental results show that the improvement in the detection phase and the innovation in the classification stage have greatly improved the overall performance of the identification system,guaranteed recognition rate while meeting the real-time requirements.