Research Of Traffic Sign Detection And Recognition Based On Multi-feature Fusion Under Driving Environment

With the development of social economy, vehicles increase with each passing day, the application of intelligent transportation system has been paid great attention by people. As the core technology of intelligent transportation system, automatic traffic sign detection and recognition get more and more scholars’ attention and research. What’s more, this core technology has a wide range of applications in driver assistance systems, unmanned vehicles and road marking maintenance. However, traffic signs appear fade, breakage, shadow, occlusion, motion blur, and the interference of similar color and shape objects in the real complex scenario. Faced with these problems, many scholars have conducted in-depth research, but the research findings are far from mature. Especially under the circumstances of the large population and the increasing popularity of private cars in our country, traffic jams and life safety problems become more and more serious. Therefore, the research of automatic traffic sign detection and recognition is of very important theoretical and practical significance.Aiming at the key technology of automatic traffic sign detection and recognition in intelligent traffic system, this dissertation mainly studies the blurry problem of traffic signs caused by vehicle acceleration or camera dithering in driving environment, the image low-level feature fusion, the color segmentation of the traffic signs and the region of interest extraction, as well as the application of support vector machine (SVM), extreme learning machine and other classifiers in traffic sign detection and recognition. The primary research work and results of this dissertation are as follows:(1) For the motionblur problem of visual images acquired by camera in driving environment, a blind image restoration algorithm based on sparse representation and Weber’s law is studied. Firstly, the shock filter is used to predict the sharp edges of the blurred image, and a multi-scale strategy is adopted to estimate the blur kernel from coarse to fine. Then, the sparse representation combined with Weber’s law as regularization constraints are introduced into the blind image restoration model. The experimental results show that the proposed algorithm can achieve superior performance, and obtain good restoration effect in the image texture. Moreover, the proposed method is consistent with the characteristics of human visual perceptionas it could reduce the boundary artifacts of the restored image.(2) For the problem that the imbalance between sample classes in traffic sign detection often leads to the weakening of the detection performance, a traffic sign detection method based on region of interest and HOG-MBLBP fusion feature is proposed. The color enhancement technology is used to segment and extract the region of interest of traffic signs from natural background according to the bright colors of traffic signs. An image low-level HOG-MBLBP fusion feature is studied, and the fusion feature is extracted from the traffic sign database.The genetic algorithm is used to optimize the selection of parameters for SVM cross validation so as to train and improve the performance of SVM classifier. Finally, the extracted HOG-MBLBP feature of regions of interest is input into the trained SVM classifier for further accurate detection and positioning, eliminating the false detection area.The proposed algorithm is tested on self-built SDU_CVPR_A traffic sign database and GTSRB database respectively, and the results verify the superiority of the proposed method.(3) In order to quickly and accurately identify the detected traffic sign, a traffic sign recognition method is proposed based on HOG-MBLBP fusion feature and extreme learning machine. Firstly, a traffic sign database of 23 categories is established according to the characteristics of traffic signs in China. Then, the HOG feature, BLBP feature, MBLBP feature and HOG-MBLBP fusion feature are extracted from the traffic sign database. The extracted features are input into ELM classifier, KNN classifier and Random Forest classifier for classifier training respectively.The experiments conducted onself-built SDU_CVPR_B traffic sign database and GTSRB database show that the fusion feature combined with ELM classifier can achieve better recognition effect.(4) In consideration of the wide application of semantic feature BoF model in image classification task, in order to better express the image, the relationship between the low-level visual features and the high-level semantic features is established. A traffic sign recognition method based on fusion feature BoF model and pyramid matching is studied. Firstly, The local invariant features are clustered by using k-means clustering method, and the dictionaries are built according to cluster centers. Then, the BoF model is represented by image histogram, and the spatial pyramid strategy is adopted to make full use of the spatial structure information of local invariant features. Finally, the SVM classifier is trained. The experimental results on self-built SDU_CVPR_B traffic sign database and GTSRB database show that the HOG-MBLBP fusion feature can achieve better classification effect,and the recognition rate of the BoF model of HOG-MBLBP fusion feature is better than ELM classification recognition of HOG-MBLBP fusion feature.In summary, some key problems involved in the traffic sign detection and recognition under driving environment are explored and studied in this dissertation, which is intended to improve the accuracy and rapidity of traffic sign detection and recognition, enrich the theoretical system of intelligent transportation system, and solve the existing traffic problems in our country to the greatest extent.