Warning Traffic Sings Recognition Based On Multi-feature Fusion

Along with the progress of society, the development of economic and the popularity of cars, traffic accidents occur frequently, most of which must be due to human factors and relate to drivers ignoring traffic signs. Warning signs as the important component of traffic signs mainly provide for warning message, which has a strong warning ability on driving behavior. Consider the shape, warning signs generally are triangles, so the effective recognition of the triangle traffic sign can timely remind drivers the change of conditions in front of the car to help drivers control the vehicle effectively. Therefore, they could improve traffic safety and minimize the loss. Nowadays, setting warning sign is the most popular control mode of warning vehicle around the world. But for various reasons, it’s difficult for drivers in the driving process to find the warning sign as early as possible. Therefore, the automatic detection and recognition of warning sign is of great significance for safe driving.As being in complex outdoor environmental conditions, warning sign’s detection and recognition are often vulnerable to the weather, light, tilt, bleaching and similar to the background, furthermore, the performance of the algorithms need to consider of not only recognition rate but also real-time. Therefore, it’s quite difficult to design recognition algorithms. This thesis has improved the existing recognition algorithm by summing up a large of research methods at home and abroad, and has designed and implemented a new type of warning sign recognition algorithm.In the detection phase, to avoid of light and the changes of weather influencing on the image color information, this thesis directly used the triangular shape characteristic of warning sign to detect the candidate areas of warning sign in the gray image using rapid radial symmetry. In order to get more details, the edge detection based on log conversion was introduced. For the purpose of reducing the computational load and further decreasing the error detection rate of warning sign, a novel noise removing approach based on HOG was presented. The proposed algorithm can overcome the impact of light and weather changes, solve the problem caused by similar background and mark adhesion, avoid of the time-consuming Hough transform, and improve the detection accuracy and reliability of the algorithm.In the recognition phase, firstly, a method of Fourier-wavelet descriptor was proposed to extract rotation invariant features which can recognize slant warning signs. Then the One-Against-One Support Vector Machines were built to identify all types of warning signs. Finally, to increase the recognition rate and reduce the false recognition rate, a multi-frame fusion method was designed, which can determine whether a warning sign is really being recognized or not.Experimental results shown that the method had a high detection and recognition rate with warning signs in car video and certain robustness with complicated scenes, including lighting and weather changes, and was a simple procedure with runs near real-time. So this thesis found a practical solution to solve the existing problems and difficulties of the warning sign recognition, and laid a good foundation for the following research on traffic signs recognition.