Construction of a traffic sign detector based on voting type co-training

In this paper, we propose a method to construct an accurate traffic sign detector with a small number of manual interactions. When using a statistical learning approach, a huge number of training samples should be prepared for constructing an accurate detector. However, in a real environment, traffic signs have various appearances, and their backgrounds vary widely, too. Therefore, it is very difficult and expensive to manually collect all possible views. Co-training is one of the semi-supervised learning techniques, that can collect training samples efficiently and automatically by using multiple classifiers. In this paper, we employ this approach for improving the accuracy of a traffic sign detector with low cost. The main contributions of this paper are the extension of the co-training method by introducing a majority voting scheme, and the introduction of this framework for improving the accuracy of traffic sign detection. By using this voting type co-training, the proposed method gathers traffic sign samples automatically and accurately, and improves the performance of the traffic sign detector. Experimental results showed that the proposed method improved the accuracy of the detector with a maximum F-measure of 0.95 from 0.72.