Research On Key Technologies Of Road Traffic Signs Detection And Recognition

Vision-based road traffic sign detection and recognition,as the key technology of intelligent transportation system,is one of the crucial channels to realize road perception in advanced driver assistance system and future comprehensive unmanned driving system.The work on the reliability and adaptability of traffic sign detection and recognition technology under complex environment(meteorological conditions)has made great progress over the last decade.However,there remain certain issues,such as the decrease of the discernability of foggy traffic sign image information in misty weather conditions,the lower detection accuracy of the small-sized traffic signs in traffic scene images,and the overcomplexity of most high-performance traffic sign recognition network models that give rise to limited applications.The detailed contents of research in the dissertation are listed as following:1)A single traffic sign image defogging method based on sky segmentation is proposed,which addresses the problem of degradation of foggy traffic sign images captured under haze meteorological conditions.Firstly,the segmentation threshold obtained from the Otsu algorithm was optimized by combining the gray features of the image,thereby the traffic sign image could be accurately segmented into sky and non-sky region;secondly,the histogram equalization algorithm was employed in the sky area,and the improved dark channel prior algorithm for optimizing transmittance is adopted in the non-sky area;finally,after re-fusion,a Gaussian filtering-based approach was untilized to further clarify the defogging image to facilitate subsequent detection and recognition.The experimental results illustrate that the method is capable to effectively restore clear and realistic images,and ensure the authenticity of traffic sign image information.2)A detection network model based on improved Yolov3 network is presented,which ameliorates the matter of the lower detection accuracy of small-size target traffic signs in traffic scene images.The residual structure of the original network 1-2-8-8-4 was simplified to 1-2-4-4-4 structure to minimize the parameters;and the scale prediction pattern of the original Yolov3 network was increased from 3 scales to 4 scales to enhance the detection perfomance of small targets;moreover,the GIo U function was introduced to diminish the loss of target location and heighten the detection accuracy.Experimental results manifest that compared with the original network,the size of the improved network model is lessened by11 percentage points;the average F value on GTSDB data set is increased by 2.98% for various size signs;compared with other methods on TT100 K data set,the experimental results on F value,model size and other indivators ulteriorly vindicate the superiority of the improved network.3)A recognition network model based on improved LeNet-5 network is proposed,which works out the issue of limited application of most high-performance but overly sophisticated traffic sign recognition network models.In the improved three-layer convolution-pooling network structure,the number and size of some convolution kernels were adjusted to fully extract the detailed information of the traffic signs and strengthen the network recognition ability;in addition,the Leaky Re LU activation function was untilized to replace the Sigmoid function in the original network to accelerate the convergence speed of the network;besides,the output layer was classified by the SVM classifier to elevate the recognition accuracy.Experimental results demonstrate that the improved lightweight network model is capable of taking into account both the recognition time and accuracy,with the recognition time is0.148 s and the recognition accuracy is over 98.3% on the GTSRB data set;the experimental results on TT100 K data set further testify that the proposed network possesses better generalization ability.