Research On Traffic Sign Detection Method Based On Convolutional Neural Network

Traffic sign detection and recognition is one of the key technologies in driverless cars,It has received extensive attention from scholars at home and abroad in recent years.With the continuous improvement of computer computing power and the rapid development of deep learning technology,object detection based on image has achieved good results.The object detection method based on convolutional neural network has broad application prospects and research value in traffic sign detection.In this paper,the traffic sign detection method based on convolutional neural network is studied.The detection model based on regression method is improved and optimized based on the characteristics of traffic signs.The traffic sign detection model based on regression method is proposed and improved by network layer connection fusionto improve the ability of the network to detect traffic signs with a small pixel ratio.The work of this paper mainly includes:(1)Three network structure design and improvement methods for detection models are proposed.Based on the network structure of YOLOv2,the 3×3 convolutional layer in the feature map is decomposed into 3×1 convolutional layer and 1×3 convolutional layer,which makes the depth of the model deepen and reduces the parameters of the model,simplifying the operation;The ResNet method has designed a variety of residual modules to mitigate the gradient disappearance and gradient explosion phenomenon that may occur in deep convolutional networks during gradient backhaul by means of cross-layer connection;replace the pooling operation in the detection network with two step size,the convolution operation uses the convolutional layer to reduce dimensionality while adding multiple 1×1 convolutional layers at the bottom of the network to extract traffic sign features.The experimental results on the GTSDB dataset show that the above three methods have a certain degree of improvement on the detection effect of the model.(2)The training method of traffic sign detection model is optimized.The data distribution of each layer is consistent by batch renormalization,and Focal Loss is used to focus on difficult classification samples to speed up the training.At the same time,based on the network structure of YOLOv2,combined with the design of network structure and the optimization of training methods and training methods,a traffic sign detection model based on regression method is proposed.On the GTSDB dataset,the detection model has a recall rate of 95.23% and cost 0.017 seconds to detect a single image,which can meet the requirements of real-time and accuracy.(3)Aiming at the fact that some of the traffic signs in the TT100 K dataset are too small in the overall image,a traffic sign detection method is proposed to detect and fuse the underlying features and top-level features of the network.The top-level features include global and location information,and the underlying features contain rich details of the image.By combining the two types of features,the ability of the model to detect traffic signs with too small pixel count is improved.The experimental results on the TT100 K dataset show that the recall rate of the traffic sign detection model based on feature fusion is 95.38%,and the average time taken to detect a single picture is 0.04 seconds,which can be achieved real-time detection of traffic signs while maintaining high detection accuracy.