Study On Detection Algorithm Of Traffic Signs From Natural Scenes

Transportation System plays an important role in mitigating the traffic pressure and reducing traffic accidents, etc. Therefore, it is getting more and more attention. As an important part of intelligent transportation systems, traffic sign detection has become one of the hot. Although research in this area has achieved initial results, with the increasingly complex and diverse road environment, the difficulties and challenges of this research is also increasing. Studying on the traffic sign detection algorithms, which can meet low miss rate, high accuracy and good real-time, is still a hot topic of great academic value and application prospects.In this thesis, traffic signs in natural environment are the object of study. And it established a database of traffic sign image of natural scenes. This thesis does a thorough research in traffic sign detection, which has large data sets in complex environment.Innovation of this thesis and the main work are as follows:1. Traffic sign image preprocessing. Pretreatment for traffic signs images under the complex background, a preprocessing algorithm based on luminance information is used to improve image quality and to reduce the impact of changes of illumination. Simulation results show that, compared with RGB-based multi-channel pre-processing algorithm, which avoids the image color distortion.2. Rough detection of traffic signs. For the color characteristics of traffic signs, traffic signs give a crude method of detecting the color information based on the method of two RGB and HSV color space segmentation result integration. Simulation results show that the method based on the detection method compared to a single color space, which can effectively reduce the illumination changes due to fading and color distortion caused by the impact.3. Accurate detection of traffic signs. Based on texture of traffic signs, the HOG+ SVM is designed to detect precisely. Simulation results show that the method based on Gabor features compared with random forest classifier combination methods has lower accuracy than it.