Traffic Sign Recognition Research Based On Neural Network

As an important study area of intelligent transportation system (ITS), traffic sign recognition have got great development, and this provides great help in our daily driving safety. Therefore, the research of traffic sign recognition system is of great significance. The fast speed of cars on highway leads to lower operability and stability of cars. Furthermore, due to a long and speeding driving, drivers’ability of judging distance will be weakened and they suffer from a chronic fatigue. The paper focuses on system accuracy and timeliness by analyzing the traffic signs on the highway.Traffic sign recognition system mainly includes image acquisition, image detection and recognition. In order to overcome the negative effects of noise on the image and achieve better filtering effect, in this paper, image equalization, image sharpening and Gamma Correction will be tested respectively. After the comparison of the several filtering methods, combined with analysis of the experiment results, the median filtering method will finally be found the best suitable method in removing the negative effects of noise.After analyzing the characteristics of different color area of HSV color space in detail, a modified color segmentation algorithm is put forward, and the experiment results indicate that the improved algorithm can get good segmentation result. Finally, the interesting region can be segmented from the image.Then, a series of morphological operations will be used in segmented region, including binary image expansion, smooth, corrosion, edge thinning, etc., to make sure that the shape of the target area is more advantageous for subsequent analysis. This paper adopts marker-based to recognize segmented region.Finally, the characteristics of image color and shape features will be analyzed respectively, and, combining with morphology, moment characteristic of the ten kinds of images will be extracted as the recognition evidence. Then the basic theory of pattern recognition will be briefly introduced, and a three layers structure and sample database which are based on BP neural network are established. On the basis of the traditional BP network classification experiments, it analyzes the limitations of traditional BP neural network. Meanwhile, through using weights determined directly, the network, which self-determination structure, is designed. Experimental results show that the proposed method has higher accuracy and better real-time.