Traffic Sign Recognition using a Generative Model of Scale Invariant Feature Descriptors

This paper presents an approach based on a generative model of Scale Invariant Feature Transform (SIFT) descriptors for the purpose of traffic sign recognition. In our work, SIFT key descriptors [1] are grouped and therefore providing rather unique descriptors of traffic signs by referencing the descriptor’s main orientation to the center of the sign and thus the single SIFT based keypoint matching is significantly improved. A careful analysis of the performance of the descriptor grouping on the real world traffic sign imagery is presented. Measures have been taken for preventing ambiguity within the voting for
different sign hypotheses. Since the approach uses SIFT features it is rather invariant to image scaling, translation, illumination changes and affine projections. The approach shows good performance on a wide range of images with different scales, in plane rotations and partial occlusions. This approach allows reliable detection of multiple traffic signs of different categories in the same image. The database used for the training of the nearest neighbor classifier consists of real world traffic signs. The performance in the detection of traffic signs of images from the IMSERV database proved to be around 80% in accuracy. The performance of this robust approach is competitive to the stateof-art approaches that are dedicated to the traffic sign recognition problem.