Technologies Of Traffic Sign Detection In Complex Scenes

As a necessary part of intelligent vehicles'vision system and a classic application of automatic target detection and recognition, researches in traffic sign detection and recognition has gained great popularity since recent years. However, because of the highly time sensitive nature of this vision task, real-time traffic sign detection and recognition still has many problems when it is used in complex scene images. Where there are such gaps, there are both challenges and possibilities, which is what makes the task so attractive to researchers, including me.In this paper, we firstly divide an image into three regions with different searching priorities based on priori experiences of signs'locations in scene images. It's a knowledge-based optimization strategy with a similar principle as the process of human vision. Experimental results show that this strategy largely improves the efficiency of detection and helps to discard noises greatly.Over a selected searching region, we propose a multi-space and multi-channel method of color detection, which is motivated by human cognitive habits of colors. To be more specific, color detection is not only based on channel H in HSV space, but also on channel R,G,B in RGB space and channel S in HSV space. Ranges of these channels, which represent priori knowledge and experience, come from statistical data of more than one hundred real scene images. Experiments indicate that this method has much better results than color detection just based on channel H.Visual Saliency Model is used in target detection simultaneously. A so-called"Two-way Fusion Mechanism", including both top-down methods and bottom-up methods, is proposed. With this mechanism, we try to imitate the detection process of human vision, in which searching subjectively and being attracted objectively happens at the same time. Experiments validate that this mechanism has competitive accuracy in sign detection in complex scene images and favorable speed for a real-time system.In sign recognition, improved Hough Transform is used for circles detection which is a basic step for further segmentation and recognition.