Abstract:

In this paper we discuss theoretical foundations and a practical realization of a real-time traffic sign detection, tracking and recognition system operating on board of a vehicle. In the proposed framework a generic detector refinement procedure based on a mean shift clustering is introduced. This technique is shown to improve the detection accuracy and reduce the number of false positives for a broad class of object detectors for which a soft response's confidence can be sensibly measured. Track of an already established candidate is maintained over time using an instance-specific tracking function that encodes the relationship between a unique feature representation of the target object and the affine distortions it is subject to. We show that this function can be learned on-the-fly via regression from random transformations applied to the image of the object in known pose. Secondly, we demonstrate its capability of reconstructing the full-face view of a sign from substantial viewangles. In the classification stage a concept of a similarity measure learned from image pairs is discussed and its realization using SimBoost, a novel version of AdaBoost algorithm, is analyzed. Suitability of the proposed method for solving multi-class traffic sign classification problems is shown experimentally for different image representations. Overall performance of the entire system is evaluated based on a prototype C++ implementation. Illustrative output generated by this demo application is provided as a supplementary material attached to this paper.