Low-Rank Matrix Recovery for Traffic Sign Recognition in Image Sequences

摘要

We consider the problem of traffic sign recognition in image sequences. In many cases, image sequences of traffic signs can be collected from consecutive videos and these images have high correlation with each other. While traditional traffic sign recognition approaches focus on how to extract better features and design more powerful classifiers, most of these methods neglected this correlation. In this paper, we introduce the low-rank matrix recovery model to exploit the correlation among images with similar appearances to enhance feature representation. By recovering the underlying low-rank matrix from the original feature matrix consists of feature vectors of image sequences, we are able to attenuate the influence of corruption, such as noise and motion blur. Experiments are conducted on GTSRB dataset to evaluate our method, and noticeable performance gain is observed by using low-rank matrix recovered from original matrix. We obtain very impressive results on several super-class accuracy while get comparable performance with state-of-the-art results on global accuracy.