Research On Traffic Sign Detection And Classification Algorithms For Intelligent Driving

As an important part of intelligent driving system, traffic sign recognition system is the key and basis of the research on intelligent driving system. It plays an important role in driving assistance, maintenance of traffic signs, autopilot and other aspects. However, some problems make the research on traffic sign recognition system face many difficulties, such as complex real traffic scenarios, light conditions, weather conditions, partial occlusion, multiple flags gathered, tilted viewing angle, the interference of similarity background color and so on. Simultaneously, real-time requirement is far from the point of maturity against the practical application.Based on summarizing the current research situation of traffic sign recognition technology, and analysis of the problems of existing methods, the work of this thesis focuses on researching traffic sign detection and traffic sign classification, as given in the following.Firstly, since some current traffic sign detection algorithms are sensitive to light and deformation, difficult to set the threshold value, so they have low accuracy, less robustness or difficulty for meeting real-time requirements, a new traffic sign detection method is proposed. which based on selective searching. It uses an improved hierarchical grouping algorithm to obtain object hypothesis regions of traffic signs, and then directly extracts HOG feature and trains these regions. It doesn't need exhaustive searching. The improved hierarchical grouping algorithm uses similarity strategy based on combined weights to merge the divided region for obtain more and better object hypothesis regions of traffic signs. Experimental results show that the proposed method achieves high accuracy ratio, robust to various adverse situations, and has a greater potential for real-time practical application.Secondly, since traffic signs classification algorithm based on convolutional neural network has too complex network structure, model training and traffic sign recognition process are time-consuming, difficulty for meeting real-time requirements, a traffic signs classification algorithm is put forward based on fast multi-scale convolutional neural network. A simpler network structure is used to automatic learn multi-scale features of traffic signs in this method, and rectified linear units(ReLu) is used to reduce time-consuming, as well as dropout policy is used to prevent over-fitting during training. Experimental results show that this method can classify the different scales of traffic signs. It not only has a higher classification accuracy, but also has a faster processing speed.Thirdly, a prototype system for traffic sign recognition is designed and implemented based on the above researches. The prototype system can carry on real-time detection and recognition for input traffic signs, and has some practical value.