Pedestrian Detection by Feature Selected Self-Similarity Features

This paper is concerned with the pedestrian detection problem. In this area, boosted decision tree (BDT) methods are highly successful and very efficient. However, to achieve the best performance, most BDT methods require a large number of input features, which make the algorithm scale poorly to largescale data. Inspired by the effectiveness of self-similarity (SS) features, we use linear discriminant analysis to select features in the SS features according to their generalized Rayleigh quotient, leading to a small number but most discriminative features. These features are called feature selected self-similarity (FSSS) features. The FSSS features are only used for the late stages of the BDT cascade, making the training and detecting much more efficient. Extensive experiments on four well-known data sets demonstrate that the FSSS features are highly effective and the trained pedestrian detector achieves state-of-the-art performance among all existing non-deep-learning methods on several benchmarks. We also compare our method with deep learning methods and show its superiority in high-quality localization and will be a good complement to deep learning methods.