A Comparative Study On Street Sign Detection

Street sign is a special type of traffic sign with text on to indicate named roads. The detection of it is of vital importance for intelligent transportation system. It is hard to directly localize the accurate position of texts on the street signs since they always appear in complex environments with many candidate words around. Thus, we decompose the task into two subtasks, the localization of the entire street signs and the recognition of texts on the detected signs. This paper mainly concerns about the first task by adopting machine learning methods. A general framework is proposed to acquire the descriptors after the analysis of several state-of-arts. Many types of features can be achieved through this framework. The detailed evaluation experimental results are given to demonstrate the performance of each kind of features using the linear support vector machine detector. Then, the second part of text recognition is discussed by analyzing the commonly used text segmentation algorithms. The main contributions of this paper are concluded as follows:(1) This paper proposes a general descriptor acquiring framework which can be casted into many types of robust features. It includes five steps:pre-processing, image transformation, block designing, local feature counting and normalization. Many different descriptors, including some state-of-arts and the new ones, can be got by considering about the information of color, gradients and texture.(2) A comparative study is conducted to evaluate the performance of several types of the descriptors which are acquired from the framework. By using the linear SVM as detectors, the classification and detection of street signs are studied. The experimental results show that the same kind of information has different description performance when utilize different combinations to extract them. Most of the descriptors achieved from the framework are robust to the problem of street sign detection.(3) A benchmarked street sign dataset is built since there exist no public ones concerning about it. The dataset include a thousand of pictures in various scenes derived from several different channels, and the exact positions of street signs in the pictures are benchmarked. The statistics of the dataset are studied including the position and scale aspects.(4) The recognition of texts on the detected street signs is conducted by adopting binaryzation methods. Several commonly used segmentation algorithms are discussed in detail.