Color fused multiple features for traffic sign recognition

Abstract

In this paper, we design the color fused multiple features to describe a traffic sign, and we further implement this description method to detect traffic signs and to classify multi-class traffic signs. At the detection stage, we utilize the GentleAdaboost classifier to separate traffic signs from the background; at the classification stage, we implement the random forest classifier to classify multi-class traffic signs. We do the extensive experiments on the popular standard traffic sign datasets: the German Traffic Sign Recognition Benchmark and the Swedish Traffic Signs Dataset. We compare eight features which include the HOG feature, the LBP feature, the color cues and their different combinations. We also compare the popular classifiers for traffic sign recognition. The experimental results demonstrate that the color fused feature achieves better classification performance than the feature without color cues, and the GentleAdaboost classifier achieves the better comprehensive performance of the binary classification, and the random forest classifier achieves the best multi-class classification accuracy.