Traiffc Sign Detection And Recognition Research Based On Visual Saliency

With the development of the social and science technology, the popularity of the motorvehicle makes the life of people more convenient, but it also causes many problems. Frequenttraffic accidents seriously affect the normal life of people. Intelligent transportation system iswidespread concerned by the governments and research institutions. As an important part ofITS system, the detection and recognition of traffic sign have become a hot spot in thetransport filed. Effective traffic sign detection and recognition can improve auxiliary andsafety of motor vehicles driving. Therefore it is important for research and application todesign a real-time and accurate traffic sign detection and recognition algorithm.In this paper, for problems in the detection and recognition of static traffic sign imagesand dynamic video sequences under natural scenes, we propose a traffic sign detection andrecognition method based on visual saliency. Our works are as follows:In the detection aspects: this paper proposes a new multiscale saliency object detectionalgorithm by combining log-gabor filter with quaternion Fourier transform phase spectrum,and computes location of traffic sign using the proposed algorithm. In addition, we realizetraffic sign automatic extraction based on visual perception, using automatic foreground andbackground brushes merged into existing interactive segmentation method to replace artificialmarker process. Experimental results show the proposed method is superior to traditionalsaliency models, and can accurately extract traffic signs from scene images.In the recognition aspects: this paper proposes a traffic sign SURF (Speeded Up RobustFeatures) local features description method, and builds traffic signs feature vector bycombining with LLC (Locality-constrained Linear Coding) method. And then, on our owndatabase including1000images of ban and non-ban traffic signs, we train a SVM (SupportVector Machine) classifier that can be used to test recognition accuracy rate. Finally, weperform experiments to verify the proposed algorithm, experimental results show that theproposed algorithm can reduce operation time, and the recognition accuracy rate is up to98%.