Method For Traffic Sign Detection Based On Faster R-CNN

Traffic sign recognition system is a major part of intelligent driving system,and it plays an integral role in many aspects,such as intelligent auxiliary driving,automatic maintenance and management of traffic signs,and so on.However,highway traffic environment in natural scenes is very complex,and a lot of problems such as weather conditions,light intensity and background interference make the relevant studies of traffic sign recognition system be confronted with many problems such as false and failure to detection.In this paper,traffic signs are recognized with the use of deep-learning object detection algorithms based on TensorFlow framework.The main work is as follows:(1)The types and character analysis of traffic signs are introduced,and the current research status and difficult problems of traffic sign detection are also reviewed and summarized respectively.Through contrastive analysis,the advantages of Deep-learning algorithms used in traffic sign detection become more obvious.The commonly used Convolutional Neural Networks models of Deep Learning are described in this paper,and the object detection algorithms are also put emphasis on,such as RCNN?Fast R-CNN and Faster R-CNN.(2)Data sets are crucial to the recognition performance of Deep-learning algorithms.Aiming at a lack of the suitable domestic traffic sign data sets,the data set based on the natural scene at home is established.Methods of Reversal and brightness processing are applied in this paper further,to solve the problem of uneven distribution of various data.Label the data set,and then convert it to TFrecord-formatted file for the sake of training and testing.(3)The system environment is set up,which is based on Windows system and GPU server platform.In order to raise network training speed and enhance the traffic sign detection performance,the structure of networks with detection models are improved and the ResNet101 is selected to do the original feature extraction of traffic signs,which determines the method of traffic-sign detection based on Faster R-CNN_ResNet-101.Optimize the training model performance by fine-tuning the model and adjusting parameters.Aiming at false detection,the method of Hard Negative Mining is presented;In order to reduce the impact of position error and false detection when small target traffic signs are in the testing process,a high level thresholding value and size of the filtering testing rectangle are set to enhance the traffic sign detection performance effectively.Finally,evaluate the traffic-sign detection model with the use of visible interface in Tensorboard.(4)The testing experiments under different illumination,different scenarios and shooting distance are designed to verify the correct and robust algorithm;beyond that,the performance of training model is examined through traffic-sign images with similar characteristics.The experimental results show that the improved traffic sign detection model can achieve a high detection rate.The traffic-sign detection model in this paper finally obtains 96.7% mAP,and can detect the traffic signs in natural outdoor scenes.