ABSTRACT
The existing traffic lights error detection methods can be classified into two classes: methods based on hardware detection (Kong, 2015; Zou, 2012), and methods based on video recognition (Diaz, 2015; Omachi, 2009; Gomez, 2014; Jie, 2015). The former is based on the measurement of internal current and voltage of traffic lights. Due to difference in internal structures of traffic lights, curves and amplitudes of current and voltage are not similar; therefore, methods based on hardware are not universal. Moreover, the countdown characters cannot be recognized by the amplitudes of the current and voltage. The methods based on video recognition always use the color, shape, and texture. However, these features are difficult to extract due to the change of lighting condition.
