Robust pedestrian detection under deformation using simple boosted features

Abstract

Many existing methods for pedestrian detection have the limited detection performance in case of deformation such as large appearance variations. To overcome this limitation, we propose a novel pedestrian detection method that uses two low-level boosted features to detect pedestrians despite the presence of deformations. One is a boosted max feature (BMF) that uses a max operation to aggregate a selected pair of features to make them invariant to deformation. Another is a boosted difference feature (BDF) that uses a difference operation between a selected pair of features to improve localization accuracy of pedestrian detection. We incorporate a spatial pyramid pool method that uses multiple sized blocks to increase the richness of boosted features in a local region and use a RealBoost method to train a tree-structured classifier for the proposed pedestrian detection method. We also apply a region-of-interest method to the detected results to remove false positives effectively. Our proposed detector achieved log-average miss rates of 19.95%, 10.39%, 36.12%, and 39.57% on the Caltech-USA, INRIA, ETH, and TUD-Brussels dataset, respectively, which are the lowest among those of all state-of-the-art pedestrian detectors.