Research On Real-Time Recognition Of Traffic Signs In Natural Environment

Nowadays, researches on pattern recognition are paid more and more attention, which makes computers or other intelligent agents play more and more significant roles in various aspects, assuring us comfortable life and convenient work situation. Especially in recent period, study of computer vision technology becomes the popular topic with the improvement in electric component of cameras and computation capability of CPUs (or GPUs), in the hope that the combination of computer and camera could assist us to make options or even decisions. This dissertation focuses on the recognition of traffic signs in pictures or video frames with high efficiency and accuracy. The research is based on the German Traffic Sign Detection and Recognition Database, and mainly works include the following aspects upon recognizing traffic signs instantaneously.Firstly, we conduct a pre-period process targeting traffic sign regions using the information of experience on the image which is read and stored in the memory. It always has been the classic issue dealing with colors. Facing various input image data, the article concentrates on the study of fast and robust ways to filter out irrelevant pixels, extract color and shape information and locate where the traffic signs might be in the natural environment. The improved circle extraction method and the proposed triangle detection method have stable optimization results and can find out the regions of interests in a short time.Furthermore, in detection period, the aim is to make the computer realize whether the region of interest is the traffic sign or not. It’s a typical two-category classification issue where SVM (Support Vector Machine) algorithm performs quite well. We are urged to select an appropriate feature descriptor and train them in the SVM classifier with little error, which can confirm the information of regions extracted and can also reassure the data resource in the following stage.Thirdly, in the part of recognition, it’s actually a multi-category classification issue. Signs in the recognition database are taken in different angle, distance and illumination situations, which provide us an available searching space to match. And we will analyze how to improve recognition rate gradually with full effort both in exploring experienced model and machine learning.Last but not least, the dissertation synthesizes the 3 stages, measures the effect and performance of some test images and videos, and approves the high efficiency of recognition. It’s a reliable method in real-time traffic sign recognition for its excellent anti-noise capability and precision rate. Especially, the time-saving shape extracting method proposed and the overall algorithm performance make it of applicable value in our traffic governing and driving assist systems.