Improving traffic sign recognition using low dimensional features

Abstract

In the recent decades, researches of the autonomous vehicle are getting popular in the computer vision society, since such vehicle is equipped with cameras for sensing the environment in helping navigation movement. Cameras give a lot of information and are low-cost device sensor rather than the other sensors which can be mounted on the vehicle. One of the visual information which can be acquired by autonomous vehicle for its navigation is traffic sign. Thus, this work addresses a traffic sign recognition framework as part of the autonomous vehicle. For recognizing the traffic sign, it is assumed that the traffic sign regions have been extracted using maximally extremal stable region (MSER). Using a heuristic rule of geometry properties, the false detections will be excluded. Furthermore, traffic sign images are classified using low dimensional features which were encoded using Adversarial Auto-encoder technique. Using this strategy, classification task can be performed using 2-dimensional features while improving the classification results over the high dimensional grayscale features. Extensive experiments were carried out over German traffic sign recognition database show that the proposed method provides reliable results.