Traffic Sign Recognition Method Based On Faster R-CNN Model

With the development of driverless technology,many intelligent technologies such as adaptive cruise control and automatic parking started to be applied to cars,bringing more convenience to people’s driving.In the process of driving,people must follow the traffic signs correctly to avoid causing traffic violations.However,human beings are always negligent and misread traffic signs or do not follow them,thus causing violations and even traffic accidents.Intelligent Traffic Sign Recognition(TSR)technology has become the direction of research for many car companies.Under natural conditions,traditional recognition methods are difficult to meet the actual needs due to the complex road environment and the influence of external factors such as light,weather and traffic sign damage.Therefore,the research of accurate and efficient traffic sign recognition technology has important application value in road safety and driverless system.This paper combines knowledge about deep learning and image processing to investigate the problem of road traffic sign recognition in natural environments.For the images collected under natural environment become blurred by light and noise,image pre-processing is used to make the extracted traffic sign feature information richer and reduce the difficulty of matching recognition.In order to solve the problems of poor robustness and slow recognition efficiency of traditional traffic sign recognition methods,an improved Faster RegionConvolutional Neural Networks(Faster R-CNN)traffic sign recognition method is proposed.The method uses Residual Network34(Res Net34)to extract traffic sign features by convolving and pooling the input image,and generates feature maps of different sizes to be fed into the Region Proposal Network(RPN)of Faster R-CNN,and then traverses the feature maps through a 3×3 sliding window.A series of Anchors are generated in the corresponding position of the original map by traversing each point of the feature map through a 3×3 sliding window,and many candidate suggestion boxes are generated according to these Anchors combined with the predicted bounding box regression parameters,and then projected onto the feature map to obtain the corresponding feature matrix,and the recognition results are obtained after Roi Pooling and a series of fully connected layer processing.And combined with difficult samples to repeat training to reduce the influence of complex environment,using Chinese Traffic Sign Detection Benchmark(CTSDB)and collecting images from actual scenes to expand the dataset,based on Open CV library,using Tensor Flow learning framework is used to load the training model for experiments and complete the training and testing of the traffic sign recognition model.After adjusting the parameters for training several times,the improved Faster R-CNN model is used to test the recognition of traffic signs under different sizes,different lighting weather,different shooting angles,different partial occlusions and similar features,etc.The recognition results are analyzed to verify the extensiveness of the method in this paper,and the results are compared and analyzed with the results of other recognition models.The innovations in this paper include the improvement of the feature extraction network,which can extract feature maps with different sizes of high and low layers to feed into the RPN network separately for classification,and the repetitive training of difficult samples.Compared with Faster R-CNN,the mean Average Precision(m AP)performance index of the method proposed in this paper is improved by 7.2%,and the average recognition time reaches 0.203s/frame,and the recognition speed meets the requirement of real-time.The results show that the traffic sign recognition method proposed in this paper has a high recognition rate with an average precision of 97% and a recall rate of 91%,while it is faster in feature extraction and recognition speed compared with traditional methods,and has good recognition results as well as robustness.