Extreme learning machine based traffic sign detection

This paper proposes a hierarchical method for traffic sign detection by employing extreme learning machine (ELM) whose infrastructure is a single-hidden-layer feedforward network. This proposed method consists of three modules: Coarse detection module, fine detection module and candidates clustering module. Histogram of oriented gradient (HOG) and color histogram are used as features of signs. This proposed method is tested on German traffic sign detection benchmark (GTSDB) data set, which has more than 900 images of German road signs covering 43 classes. The architecture of this proposed method is simple and it has strong extensionality. Experimental results have shown that this proposed method achieves 98.60% in terms of area under curve (AUC) for all categories of traffic signs in the dataset.