Virus-evolutionary genetic algorithm based selective ensemble classifier for pedestrian detection

In pedestrian detection system, it is critical to determine whether a candidate region contains a pedestrian both quickly and reliably. Therefore, an efficient classifier must be designed. In general, a well-organized assembly classifier outperforms than single classifiers. For pedestrian detection, due to the complexity of scene and vast number of candidate regions, an efficient ensemble method is needed. In this paper, we propose a virus evolutionary genetic algorithm (VEGA) based selective ensemble classifier for pedestrian detection system, in which only part of the trained learners are selected and participate the majority voting for the detection. Component learners are trained with diversity and then VEGA is employed to optimize the selection of component learners. Moreover, a time-spending factor is added to the fitness function so as to balance the detection rate and detection speed. Experiments show that, comparing with typical non-selective Bagging and GA-based selective ensemble method, the VEGAbased selective ensemble gets better performance not only in detecting accuracy but also in detection speed.