Robust and real-time traffic light recognition based on hierarchical vision architecture

`Circle´ and `arrow´ traffic lights are both common at intersections in urban road environment. However, existing purely vision based systems are only focus on either `circle´ or `arrow´ traffic light recognition, which limits their real-world application. In this paper, A novel robust and real-time traffic light recognition system based on hierarchical vision architecture is carefully designed. The system first learns a joint color space filter in normalized RGB and HSI color space to automatically select the traffic light candidate regions with connected component analysis, then detects traffic lights in the image neglecting the direction information with Multi-layer Histogram of Oriented Gradients (MHOG) feature, at last, it determines the direction information with the traditional Histogram of Oriented Gradients (HOG) feature extracted from the light-emitting unit and linear SVM classifiers. Experimental results show that the whole system can provide high recognition rate, and it can provide robust and real-time decision support for the intelligent vehicle.