Road Traffic Sign Detection And Recognition Based On Deep Convolutional Neural Network

The rapid development of autonomous driving leads to the increasing demand for road traffic sign detection and recognition,and the rapid development of deep convolutional neural network has made remarkable progress in road traffic sign detection and recognition system.However,there are still some problems such as missed detection,false detection,inaccurate positioning and low confidence in cloudy days,evening or backlight.In this thesis,in order to solve the problem of road traffic sign detection and recognition under complex lighting conditions,the following work has been done:（1）Aiming at the lack of suitable data sets of complex illumination road traffic signs in China,the Chinese road traffic sign data set ZCTSDB under the real road traffic scene is constructed,and seven different image amplification strategies are used to enhance its data,so as to improve the balance of road traffic sign categories in the training set.Subsequently,the ZCTSDB is labeled and transformed into a COCO mode convenient for training and testing.Experiments show that the seven training set data augmentation strategies adopted in this thesis can effectively improve the performance of road traffic sign detection and recognition.Compared with ZCTSDB-original,the average accuracy of ZCTSDB-augmentation target detection and instance segmentation increased by 1.288%and 1.206%,respectively.（2）Aiming at the problems of low brightness and low significance of traffic signs in low illumination traffic scenes,a new illumination image enhancement algorithm is proposed in this thesis.Firstly,RGB color space is transformed into HSV color space,and the illumination component is obtained by multi-scale Gaussian filtering for V channel.On the one hand,the V-channel reflection component is normalized to obtain the corrected V-channel image.On the other hand,another corrected V-channel image is obtained by using the improved adaptive gamma function.The two corrected V-channel images are fused with weights to obtain a new V-channel image,which is combined with H and S components and converted into RGB color space to obtain a new image.Finally,the new image and the original image are reconstructed by gradient formula and parameter adjustment.Experiments show that the illumination image enhancement algorithm proposed in this thesis can effectively improve the image brightness and image definition in various low illumination traffic scenes,and effectively improve the performance of road traffic sign detection and recognition without image distortion.Compared with Mask RCNN,the average accuracy of combined with illumination image enhancement and Mask RCNN algorithm object detection and instance segmentation improved by 2.213%and 2.774%respectively.（3）Aiming at the problems of missing detection,false detection,inaccurate positioning and low confidence in road traffic sign detection and recognition in complex illumination environment,a road traffic sign detection and recognition algorithm under complex illumination is proposed.Firstly,the real-time illumination intensity judgment algorithm is used to judge the illumination intensity of the image;Secondly,the illumination image enhancement algorithm is used to adjust the brightness and contrast of the image,improve the significance of road traffic signs,enhance the details of road traffic signs,and reduce the missed detection and false detection rate;Finally,mask RCNN detection image based on Res Next-101-FPN improves the performance of road traffic sign detection and recognition.In this thesis,an improved road traffic sign detection and recognition algorithm based on deep convolution neural network is proposed,which is trained and tested on ZCTSDB data set.Experiments show that the algorithm proposed has good recognition effect on 43 types of road traffic signs on ZCTSDB test set.The average accuracy of target detection8)⁽（7（7） is 72.047,and the average accuracy of instance segmentation8）^seg is 73.824.Experimental results show that the proposed algorithm can effectively improve the detection and recognition accuracy of road traffic signs in complex lighting traffic scenes.