Beyond Sliding Windows: Saliency Prior Based Random Partition for Fast Pedestrian Detection

Recently many powerful complicated features have been used for pedestrian detection successfully but they are not fit for real applications because of heavily consuming time caused by production of complicate feature extraction and millions of candidate object probing. The formal is critical for pedestrian detection, so for solving this problem, effective region proposal strategy was proposed. Such approaches generate candidate regions either by segmentation or by shape classification, and they still generate several thousand regions each image, that is too many for fast pedestrian detection. In this paper, a novel search strategy, saliency prior based random partition, is proposed to generate nearly two hundred regions and consume less time than selective search at the same recall. And we prefer the Deformable Part Model [8], one of the most popular object detectors, as the pedestrian detector. At last, we combine the salient prior and the part based detector by Bayesian inference. Experiment results on INRIA person dataset and Caltech person dataset have demonstrated that our approach has outperformed the selective search method.