Traffic Sign Detection And Recognition Based On Multiresolution Convolutional Neural Networks

Setting traffic signs on motor vehicle railway plays an important role in adjusting traffic flow and improving the road capacity. It also makes drivers be directed in the current railway conditions and take corresponding measures in advance, which reduces the probability of traffic accidents. With the increasing of the number of vehicles and more complicated railway conditions in recent years, traffic safety problem has drawn a heated discussion. Through integrating advanced science and technology, Intelligent Transportation Systems(ITS) builds a safety driving system that controls the vehicles automatically.As one of the most significant functions in traffic driver assistance system, Traffic sign recognition plays an important role in this issue. Years of research has been made by governments and vehicle manufacturers. However, there are still many that need to be resolved because of the complexity of the real traffic conditions. Therefore, traffic sign detection and recognition are introduced as the study topic in this dissertation, convolutional neural network algorithm is applied to this research.In order to improve the speed and accuracy of traffic sign detections in scene image, this dissertation proposes an asymmetric convolutional neural network(CNN) algorithm. This algorithm firstly extracts ROI areas from original image by using traditional color conversion and shape matching methods, and then classify the preceding results. Simulation results of the German traffic sign detection benchmark data sets show that the proposed algorithm can overcome the harmful effect such as under-size image, adverse weather, color fading and so on, which proves the algorithm’s effectiveness and robustness.To solve the problem of long training time in traffic sign recognization, a multi-resolution convolutional neural network algorithm is proposed. On the basis of CNN theory, this algorithm uses two different branches to train network and extracts features, then makes classification. By preprocessing the image into a high-revolution and a low-revolution image data sets, these two data sets are servers as the inputs of the two branches respectively. The high-revolution branch extracts the global contour feature where the low-revolution branch extracts the local detail feature at the same time. After the integration of the features which are extracted by the two branches using the whole-connection layer, the classification and recognization are made. Simulations results of the GTSRB data sets show that the proposed algorithm reduces the training time dramatically on the basis of high recognization accuracy, which balances the accuracy and real-time performance.